© Maha Ben Ali, 2018

Approches intégrées de gestion de la demande dans l'industrie de bois d'oeuvre

Thèse

Maha Ben Ali

Doctorat en génie mécanique

Philosophiæ doctor (Ph. D.)

Québec, Canada

Approches intégrées de gestion de la demande dansl’industrie du bois d’œuvre

Thèse

Maha Ben Ali

Sous la direction de:

Sophie D’Amours, directrice de rechercheJonathan Gaudeault, codirecteur de recherche

Résumé

Les entreprises du bois d’œuvre ont besoin d’une part de synchroniser les activités de pro-duction, d’approvisionnement et de ventes, et d’autre part, de maximiser les profits face àune demande hétérogène et saisonnière. L’objectif de cette thèse de doctorat est de dévelop-per et d’évaluer de nouvelles approches intégrées de gestion de la demande dans un contextede capacité limitée, afin de mieux orienter ces entreprises de façon à maximiser les profits età améliorer la satisfaction des clients les plus prioritaires.

Le cas d’étude, inspiré de la réalité des entreprises du bois d’œuvre québécoises, considèredifférents clients hétérogènes, des processus de production divergents et plusieurs usinesœuvrant dans un mode de fabrication pour les stocks, où les plans d’approvisionnement, deproduction et de ventes sont pilotés par les prévisions de la demande et des prix.

Dans cette thèse, on commence par définir un cadre décisionnel mutiniveau afin de sup-porter les décisions d’allocation prises aux niveaux tactique et opérationnel, ainsi que lespromesses de livraison conclues en temps réel. En particulier, on propose un processus degestion de la demande intégrant la planification des ventes et des opérations (S&OP) avecla promesse de livraison basée sur les concepts de gestion des revenus : une nouvelle for-mulation mathématique, intégrant un modèle S&OP adapté à l’industrie du bois d’œuvreet un modèle de promesse de livraison utilisant des limites de réservation imbriquées, estfournie. Cette formulation offre la possibilité de changer les décisions d’allocation en tempsréel tant que les commandes fermes ne sont pas expédiées. Une plateforme d’optimisationet de simulation en horizon roulant est développée afin d’évaluer la valeur de l’intégrationde la planification des ventes et des opérations et de la gestion des revenus. Les résultats desimulation ont démontré la capacité d’un processus intégrant le S&OP et la gestion des reve-nus, dans un contexte de capacité limitée et face à une demande hétérogène et saisonnière,à réaliser des meilleurs profits et à mieux satisfaire les clients prioritaires que les processusconventionnels de gestion de la demande.

La plateforme d’optimisation et de simulation développée est utilisée, dans une secondeétape, pour étudier la performance de différents processus intégrés de gestion de la demandeface à différentes situations du marché. En effet, on compare différentes séquences d’arrivéedes commandes et deux approches de promesse de livraison. A cette fin, on adopte une nou-

iii

velle procédure de planification et d’analyse des expériences dans un contexte de gestionde la chaîne d’approvisionnement : un plan de remplissage d’espace est utilisé pour définirdes scénarios variés du marché et des métamodèles de krigeage sont générés pour analyserles résultats. L’analyse des résultats a mis en évidence l’amélioration potentielle de perfor-mance qu’on peut atteindre en utilisant les concepts de gestion des revenus et a démontrél’impact de la séquence d’arrivée des commandes sur le profit annuel et le niveau de satis-faction des clients prioritaires. Les implications managériales qui découlent de cette analysesont également présentées.

Dans une troisième étape, on analyse l’effet de la substitution et de certains concepts de ges-tion des revenus. En particulier, on investigue l’intérêt d’offrir à certains clients privilégiésun produit de qualité supérieure au prix du produit original demandé (soit l’équivalent d’unsur-classement pour les entreprises de service), ce qui est une pratique assez commune dansl’industrie du bois d’œuvre. A cette fin, on introduit la dimension produit dans le modèlede promesse de livraison basée sur les concepts de gestion des revenus et on mène une si-mulation en horizon roulant afin de comparer différentes approches intégrées de promessede livraison. Les résultats de simulation soulignent l’efficacité d’une approche de promessede livraison intégrant le sur-classement et les concepts de gestion des revenus, comparéeaux pratiques communes utilisées pour satisfaire la demande dans un contexte de capacitélimitée. En effet, le sur-classement s’avère significativement bénéfique s’il est associé auxconcepts de gestion des revenus, vu que l’utilisation de limites de réservation empêche deproposer un sur-classement si le produit en question a été alloué à des commandes plusrentables.

iv

Abstract

This thesis addresses the need of softwood lumber firms operating in a supply-constrainedenvironment and facing heterogeneous and seasonal market, to synchronize between thedifferent business units of supply chain and to maximize profits. The objective is to de-velop and to evaluate new integrated demand management approaches for limited capacitycontexts in a way to maximize profits and enhance the service level offered to high-prioritycustomers.

Our case study, inspired from softwood lumber manufacturers located in Eastern Canada,considers heterogeneous customers, divergent production processes and several mills con-sidered as an MTS environment since operations and sales plans are driven by forecasts.

In this thesis, we first define a multilevel decision framework in order to support medium-term, short-term and real-time sales decisions. We propose a demand management pro-cess integrating sales and operations planning (S&OP) and revenue management (RM) con-cepts : we present a new mathematical formulation integrating an S&OP network modelin the softwood lumber industry and an order promising model using nested booking lim-its. This formulation offers the possibility of changing decisions of how confirmed ordershave to be fulfilled as late as possible, which we called order reassignment. Consideringcurrent demand management practices and existing IT-systems, we developed a simulation-optimization platform in order to evaluate the demand management process performancethe benefits of integrating S&OP and RM concepts in various scenarios. Simulation resultsprovide evidence of the value of integrating RM and S&OP and show that we can offer bet-ter service level to high-priority customers and higher profit margin compared to commondemand management practices.

The simulation-optimization platform is used, in a second step, to investigate how an inte-grated demand management process, that can be configured differently, can perform facingvarious order arrival sequences and market disturbances. For this purpose, we use rela-tively novel techniques – a space-filling design and Kriging metamodeling – in supply chainsettings to address the impact of decision and environmental factors on the performance ofthe integrated demand management process. The simulation results affirm the use of nestedbooking limits can be a powerful tool to maximize revenues facing different environmental

v

conditions. We also show how order arrival sequence can play a relevant role, especiallywith a high customer heterogeneity. In addition, as motivated by an industrial problem, wediscuss the potential implications of the analysis presented for firms operating in supply-constrained environments, such as Canadian softwood firms.

As a third step, we investigate the benefits of integrating revenue management and prod-uct substitution in a manufacturing context. We particularly examine the situation when ahigher quality substitute is provided at the original product’s price, which is called an up-grade. Upgrading is a common demand fulfillment practice in the Canadian softwood lum-ber industry. Thus, we generalize the order promising model using nested booking limitsand we add a product dimension to enable product substitution. Then, we conduct a rollinghorizon simulation in order to compare different demand fulfillment approaches. The simu-lation results demonstrate that integrating RM and upgrading achieves better performancethan common demand fulfillment approaches in a limited capacity context. The value ofupgrading is more significant when integrated with RM concepts since the use of nestedbooking limits prevents from doing unprofitable upgrades. Thus, inventories are preservedfor future profitable orders.

vi

Table des matières

Résumé iii

Abstract v

Table des matières vii

Liste des tableaux ix

Liste des figures x

Liste des abbréviations xii

Remerciements xv

Avant-propos xvii

Introduction générale 11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.2 Problématique de recherche . . . . . . . . . . . . . . . . . . . . . . . . . . . 21.3 Concepts préliminaires . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71.4 Contributions de recherche et structure de la thèse . . . . . . . . . . . . . . 191.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2 Intégration de la gestion des revenus et de la planification des ventes et desopérations dans un environnement de fabrication pour les stocks : Cas del’industrie du bois d’œuvre 25Résumé . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282.2 Problem statement and related literature . . . . . . . . . . . . . . . . . . . . 292.3 Proposed demand management process . . . . . . . . . . . . . . . . . . . . 342.4 Model formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372.5 Data generation and experiments . . . . . . . . . . . . . . . . . . . . . . . . 472.6 Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492.7 Conclusion and future research . . . . . . . . . . . . . . . . . . . . . . . . . 53

A Data generation 55

B Simulation algorithm 57

vii

3 Configuration et évaluation d’un processus intégré de gestion de la demandevia un plan de remplissage d’espace et la technique de krigeage 59Résumé . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 613.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 623.2 Related literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643.3 Industrial context and case study . . . . . . . . . . . . . . . . . . . . . . . . 663.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 683.5 Data generation and analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 733.6 Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 753.7 Conclusion and further research opportunities . . . . . . . . . . . . . . . . 82

C Analyzed factors in sales and operations planning (S&OP) literature 84

D Analyzed factors in literature about revenue management(RM) in manufac-turing 86

E Recent literature of conventional DOE for simulation systems in supply chainsettings 88

F Supplementary Materials 89S1 Steps for data generation and analysis . . . . . . . . . . . . . . . . . . . . . 89S2 Normality tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93S3 ANOVA tables for YPM and HPFR . . . . . . . . . . . . . . . . . . . . . . . 95

4 Simulation d’une approche intégrée de gestion des revenus pour un systèmede coproduction avec substitution de produits 98Résumé . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1004.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1014.2 Related literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1024.3 Mathematical formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1034.4 Application to softwood lumber case study . . . . . . . . . . . . . . . . . . 1104.5 Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1144.6 Conclusion and future work . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

G Additional demand fulfillment approaches 118

Conclusion générale et perspectives 1205.1 Conclusion générale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1205.2 Perspectives de recherche . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

Bibliographie 128

viii

Liste des tableaux

2.1 Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392.2 Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392.2 Parameters (continued) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402.3 Decision variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402.4 Additional notation for the order promising model . . . . . . . . . . . . . . . 442.5 Scope of the simulated case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 472.6 Demand scenarios considered for data generation . . . . . . . . . . . . . . . . 482.7 Simulated demand management processes . . . . . . . . . . . . . . . . . . . . 492.8 Benefits of integrating S&OP and NBL compared to process A-FCFS . . . . . 50

3.1 Factors and their associated categories/ranges . . . . . . . . . . . . . . . . . . 70

C.1 S&OP literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

D.1 RM in manufacturing literature . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

E.1 Conventional DOE for simulation systems in supply chain settings . . . . . . 88

S1.1 Environmental scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

4.2 Parameters and decision variables . . . . . . . . . . . . . . . . . . . . . . . . . 1034.2 Parameters (continued) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1044.1 Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1054.3 Scope of the simulated case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1104.4 Customer segments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1114.5 Different consumption models . . . . . . . . . . . . . . . . . . . . . . . . . . . 1134.6 Simulation scenarios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

G.1 Additionnal parameters and decision variables . . . . . . . . . . . . . . . . . . 118

ix

Liste des figures

1.1 Réseau d’approvisionnement d’une entreprise du bois d’œuvre . . . . . . . . 31.2 Différents types de clients du bois d’œuvre . . . . . . . . . . . . . . . . . . . . 41.3 Exemple de plans de coupe. Adapté de Vila, Martel, and Beauregard (2006) . 41.4 Structure et contributions de la thèse . . . . . . . . . . . . . . . . . . . . . . . . 61.5 La position de la gestion de la demande dans la gestion de la chaîne d’appro-

visionnement. Adapté de Mentzer, Myers, and Stank (2007) . . . . . . . . . . . 71.6 La coordination via la planification des ventes et des opérations. Adapté de

Mentzer, Myers, and Stank (2007) . . . . . . . . . . . . . . . . . . . . . . . . . . 91.7 Différents niveaux de planification. Adapté de Feng, D' Amours, and Beaure-

gard (2008) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121.8 Point de découplage selon le type d’environnement de production. Adapté de

Fleischmann and Meyr (2003) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131.9 Comparaison d’un plan factoriel à un plan de remplissage d’espace pour le

cas de deux facteurs. Adapté de Soderborg (2009) . . . . . . . . . . . . . . . . 191.10 Éléments de recherche du premier article . . . . . . . . . . . . . . . . . . . . . 201.11 Éléments de recherche du deuxième article . . . . . . . . . . . . . . . . . . . . 211.12 Éléments de recherche du troisième article . . . . . . . . . . . . . . . . . . . . . 23

2.1 Available forecasts for short-term and medium-term horizons . . . . . . . . . 352.2 Proposed demand management process . . . . . . . . . . . . . . . . . . . . . . 362.3 Supply network of a multi-site softwood company . . . . . . . . . . . . . . . . 382.4 Allocations assignments to quantity requested by segment g

′for due date t

′

(example where all transportation delays are set to zero) . . . . . . . . . . . . 452.5 Yearly profit margin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

B.1 Simulation algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

3.1 The integrated demand management process (IDMP) proposed by Ben Aliet al. (2014) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

3.2 The case study : A supply network of a multi-site softwood company . . . . . 673.3 Procedure for designing and analyzing experiments (adapted from Montgo-

mery (2009)) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 693.4 Performance measures and factors . . . . . . . . . . . . . . . . . . . . . . . . . 703.5 Experimental design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 723.6 Performance measures for different decision factor combinations . . . . . . . 76

x

3.7 Response surfaces for the effects of demand intensity (I) and demand forecasterror (E) on the yearly profit margin (YPM) for FCFS and NBL approaches(customer heterogeneity H=10%, coefficient of variation V=0.5 and randomarrival sequence) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

3.8 Variation of sales and inventories over a year considering different demandforecast errors (E=0% and E=20%), NBL approach, random arrival sequence,I=1.5, H=10% and V=0.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

3.9 Variation of sales and inventories over a year considering different demandforecast errors (E=0% and E=20%), FCFS approach, random arrival sequence,I=1.5, H=10% and V=0.5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

3.10 JMP profiler tool for I=1.5, E=0, H=10% and V=0.5 . . . . . . . . . . . . . . . . 79

S1.1 Steps for data generation and analysis (1) . . . . . . . . . . . . . . . . . . . . . 91S1.2 Steps for data generation and analysis (2) . . . . . . . . . . . . . . . . . . . . . 92S2.1 Normality tests for FCFS approach and ASC sequence . . . . . . . . . . . . . 93S2.2 Normality tests for FCFS approach and RAND sequence . . . . . . . . . . . . 93S2.3 Normality tests for FCFS approach and DESC sequence . . . . . . . . . . . . 94S2.4 Normality tests for NBL approach and ASC sequence . . . . . . . . . . . . . . 94S2.5 Normality tests for NBL approach and RAND sequence . . . . . . . . . . . . 95S2.6 Normality tests for NBL approach and DESC sequence . . . . . . . . . . . . . 95S3.1 ANOVA for FCFS approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96S3.2 ANOVA for NBL approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

4.1 The integrated demand management process proposed by Ben Ali et al. (2014) 1024.2 S&OP model and different consumption models . . . . . . . . . . . . . . . . . 1034.3 Supply network of a multi-site softwood company . . . . . . . . . . . . . . . . 1104.4 Profit of the consumption models in the base case scenario . . . . . . . . . . . 1154.5 Profit of the RM and FCFS approaches with/without upgrading in different

scenarios and the percentage of the profit achieved by each approach compa-red to the GO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

xi

Liste des abbréviations

ANOVA Analysis of varianceATO Assemblate To OrderATP Available To PromiseBL Booking LimitsCAC Central Canadian marketCAE Eastern Canadian marketCRM Customer Relationship ManagementCTO Configure To OrderDOE Design Of ExperimentsFCFS First-Come First-ServedHP High-PriorityHPFR High-Priority Fill RateLHD Latin Hypercube DesignLP Linear ProgramMMFBM Million Board-Feet MeasureMTO Make To OrderMTS Make To StockNBL Nested Booking LimitsRM Revenue ManagementS&OP Sales and Operations PlanningUPG UpgradingUS Northeastern American marketYPM Yearly Profit MarginYS Yearly Sales

xii

À mes parents, À mon cher Nidhal,À mes princesses Oswa et Aya

xiii

Quand je tends vers un but,J’enfourche l’espérance et oublietoute prudence.Celui qui n’aime pas grimper lesmontagnes,Vivra éternellement entre lesfossés.

Abou Al Kacem Alchabbi (1933)

xiv

Remerciements

Je tiens à exprimer ma gratitude envers ma directrice de recherche Prof. Sophie D’Amourspour ses encouragements et son soutien académique et personnel durant cette thèse. Elle aété toujours à l’écoute et prête à donner conseil pour le bien de ma carrière. Je la remercieégalement de m’avoir transmise son ouverture d’esprit et sa passion pour la gestion de lachaîne d’approvisionnement.

Ce travail n’aurait jamais pu voir le jour sans Prof. Jonathan Gaudreault qui a codirigé cettethèse. C’est grâce à ses précieux conseils, son souci du détail et son expertise en matière desimulation et de design d’expériences qu’on a réussi à réaliser un travail de qualité. Je leremercie de m’avoir guidée tout au long de ce doctorat.

J’ai eu également la chance de travailler en collaboration avec Prof. Marc-André Carle. Je leremercie infiniment pour son implication et son support tout le long de cette thèse. Sa grandeconnaissance de l’industrie forestière a grandement alimenté nos réflexions.

J’aimerais aussi remercier tout spécialement Philippe Marier, professionnel de recherche auconsortium de recherche FORAC, de m’avoir guidée surtout au début du projet. Son exper-tise, son professionnalisme et son dévouement ont été fortement appréciés.

Mes remerciements vont également aux experts, M. Michel Vincent (Conseil de l’industrieforestière du Québec) et M. François Robichaud (Forest Economic Advisors), pour m’avoirsi bien accueillie et tant apprise sur le marché du bois d’œuvre canadien.

Je remercie également Prof. Alexandre Dolgui et Prof. Nadia Lehoux pour avoir accepté dedonner de leur temps précieux pour évaluer ce travail et faire partie des membres du jury.

Je voudrais aussi exprimer ma reconnaissance pour le consortium FORAC pour son sup-port financier et académique durant la thèse. Un grand merci à tous les membres de FO-RAC, en particulier Catherine Lévesque pour son support administratif et les professeursMikael Rönnqvist et Tasseda Boukherroub pour leurs conseils. Je remercie également Riadh,Alexandre, Rémy, Sébastien et Edith pour leur support.

Je tiens aussi à remercier tous mes collègues au FORAC pour les différents moments partagéstout au long de mon parcours universitaire. Je remercie particulièrement Foroogh, Chourouk

xv

et Raja pour tous les échanges effervescents qu’on a eus, ainsi que Jean, Ludwig et Julie pourleurs encouragements.

Un grand merci à toute ma famille de m’avoir supportée dans ce parcours, à mes parentsHela et Mohsen, à ma soeur Saba et à ma belle famille, Khaouira, Ali, Nawel, Nefaa et Ma-rouene. Je remercie également tous les membres de ma famille avec lesquels j’ai partagé cetobjectif commun de mener un projet de doctorat. Intissar, Hamza et Wiem, bonne chancepour la suite !

Ce travail n’aurait jamais pu voir le jour sans mon cher Nidhal, qui m’a encouragé à en-treprendre ce programme de doctorat et à commencer une nouvelle vie au Canada. Je teremercie pour ton support et ta patience. Finalement, je remercie mes filles Oswa et Aya quiont grandi avec ce projet de doctorat. Je vous promets mes chères de me rattraper pour toutle temps où je n’étais pas présente avec vous !

xvi

Avant-propos

Ce projet, intitulé "Approches intégrées de gestion de la demande dans l’industrie du boisd’œuvre", est réalisé dans le but d’obtenir le grade de Doctorat en Génie Mécanique (Ph.D.)de l’Université Laval. Il a été effectué, au sein du consortium de recherche FORAC, sousla direction du Pr. Sophie D’Amours et sous la codirection de Pr. Jonathan Gaudreault. Ceprojet a été financé par le consortium FORAC et par deux bourses d’excellence obtenues desorganismes subventionnaires FRQNT et CRSNG.

Cette thèse est rédigée selon le principe d’insertion d’articles. Elle se compose de trois articlesqui ont été tous coécrits avec Pr. Sophie D’Amours, Pr. Jonathan Gaudreault et Pr. Marc-André Carle. Pour chacun des articles présentés, j’ai agi à titre de chercheur principal dansl’identification de la problématique, l’implantation des différents modèles d’optimisation, laréalisation des expériences, l’analyse des résultats, la rédaction du manuscrit, ainsi que larévision des versions soumises aux journaux et aux conférences.

Le premier article, intitulé "Integrating Revenue Management and Sales and OperationsPlanning in a Make-To-Stock environment : Softwood lumber case study", a pour auteursMaha Ben Ali, Sophie D’Amours, Jonathan Gaudreault et Marc-André Carle. Il a été sou-mis au journal "INFOR : Information Systems and Operational Research" en janvier 2018. Laversion présentée dans cette thèse est identique à la version soumise.

Le deuxième article, intitulé "Configuration and evaluation of an integrated demand ma-nagement process using a space-filling design and Kriging metamodeling", a pour auteursMaha Ben Ali, Sophie D’Amours, Jonathan Gaudreault et Marc-André Carle. Il est publiédans le journal " Operations Research Perspectives", Volume 5, Janvier 2018, Pages 45-58. Ila gagné le prix "4th David Martell Student Paper Prize in Forestry" de la société canadiennede recherche opérationnelle. La version présentée est identique à la version publiée.

Le troisième article, intitulé " Simulating an integrated revenue management approach ina coproduction system with product substitution", a pour auteurs Maha Ben Ali, SophieD’Amours, Jonathan Gaudreault et Marc-André Carle. Il a été accepté dans la conférence"Winter Simulation Conference 2018" en juin 2018. Une version étendue de cet article estprésentée dans cette thèse.

xvii

Introduction générale

1.1 Introduction

L’industrie du bois d’œuvre est au cœur des activités du secteur forestier canadien puis-qu’elle fait essentiellement appel à des intrants disponibles au Canada. En effet, la grandemajorité du bois d’œuvre est produite à partir d’arbres résineux comme l’épinette, le pin,le sapin et le cèdre, qui composent l’essentiel des forêts canadiennes. Cette industrie créedes milliers d’emplois directs et des bénéfices signifiants supportant des emplois indirects,en particulier dans les régions rurales et éloignées. Elle est également très présente sur lesmarchés d’exportations et réalise 20% des exportations totales du secteur forestier 1.

L’industrie canadienne du bois d’œuvre fait face ces dernières années à plusieurs défis com-merciaux et économiques, notamment les fortes pressions commerciales imposées par lesÉtats-Unis représentant le principal marché d’exportation (soit l’imposition de droits d’anti-dumping 2 et compensateurs 3 par les États-Unis sur les produits forestiers canadiens), l’aug-mentation des coûts d’opérations, ainsi que la compétition à faible coût des produits prove-nant des pays émergents en Asie et en Amérique du Sud. Un tel environnement agressifnécessite des processus intégrés qui permettent de supporter la prise de décision, non en sebasant sur la production de masse mais sur l’utilisation efficace des ressources.

Outre les stratégies de réduction des coûts d’opérations, les entreprises du bois d’œuvre ca-nadiennes doivent s’orienter plus vers des approches de gestion de la demande davantagecentrées sur le client. Étant un processus complexe dépendant de plusieurs activités exécu-tées par différentes fonctions de l’entreprise et à différents niveaux de planification, la ges-tion de la demande requiert une intégration tout le long de la chaîne d’approvisionnement,une segmentation raffinée de la clientèle et une capacité à aligner la production affectée parla disponibilité cyclique des ressources (la matière première et la main d’œuvre), face à unedemande variable.

1. Ressources naturelles Canada, Demandes en produits forestiers, http://www.rncan.gc.ca/forets/

industrie/demandes/13318, consulté le 4 mai 20182. Prélèvement sur une marchandise importée afin de protéger l’industrie nationale contre les dommages

causés par la vente de biens à des prix inférieurs à ceux pratiqués sur le marché national (Dufour 2007)3. Prélèvement sur une marchandise importée qui a pour but de protéger une industrie nationale d’un

dommage causé par des importations subventionnées (Dufour 2007)

1

Depuis plusieurs années, les praticiens et les chercheurs accordent une importance crois-sante au développement de méthodologies et d’outils d’aide à la décision pour guider lescompagnies à mieux synchroniser les activités de production, d’approvisionnement et deventes afin de saisir les opportunités du marché et gérer efficacement l’offre et la demande.Cette thèse de doctorat s’inscrit dans cette thématique de recherche et a comme objectif dedévelopper et d’évaluer de nouvelles approches intégrées de gestion de la demande afinde mieux orienter les entreprises de façon à maximiser les profits, tout en considérant lesréalités de l’industrie canadienne du bois d’œuvre.

Dans ce qui suit, nous détaillerons en premier lieu la problématique de recherche. En deuxièmelieu, nous présenterons les différents concepts abordés dans cette thèse. Nous exhiberons parla suite la méthodologie utilisée dans les différentes contributions de la thèse, ainsi que lastructure de la thèse.

1.2 Problématique de recherche

Les responsables des ventes et des opérations disposent de différents types d’informationsur la demande et sur les prix afin de planifier les actions présentes et futures. En particulier,dans un contexte de fabrication pour les stocks (Make-To-Stock MTS), tel est le cas pour laplupart des entreprises du bois d’œuvre, on se base sur des prévisions agrégées (définiespar exemple par marché, par famille de produits et par mois) afin de générer des plans àmoyen terme. Les processus de planification tactique, tels que la planification des ventes etdes opérations (Sales and Operations Planning S&OP), permettent de décider des actionsfutures d’approvisionnement, de production, de ventes et de transport en considérant lanature cyclique de l’industrie du bois d’œuvre. En outre, des prévisions moins agrégées(définies par exemple par produit, par client et par semaine) pour un horizon à court termepeuvent être disponibles périodiquement.

Toutefois, étant basés sur des prévisions et des informations agrégées sur la demande, lesplans tactiques (définis sur un horizon à moyen terme) et opérationnels (définis sur un hori-zon à court terme) ne suffisent pas pour supporter la prise de décision à temps réel. En par-ticulier, la promesse de livraison constitue un problème à temps réel auquel les entreprisesdu bois d’œuvre font face quotidiennement. Une entreprise de moyenne ou grande taille a àtraiter entre 1000 et 3000 commandes par semaine de façon individuelle et quasi-instantanée,vu que les clients dans un contexte MTS s’attendent à une réponse rapide (Quante, Fleisch-mann, and Meyr 2009). Or, ces décisions de promesse de livraison en temps réel peuventaffecter la capacité à satisfaire les demandes futures.

Durant ces dernières années, les entreprises dans différents secteurs, y compris le secteurforestier, ont bénéficié de l’utilisation des systèmes de planification avancée en tant qu’outilsd’aide à la décision. De plus, les entreprises sont amenées à intégrer de nouvelles approches

2

de gestion de la demande, telles que les approches basées sur les concepts de gestion desrevenus (Revenue Management RM), afin d’acquérir un avantage concurrentiel et de de-meurer viables face aux différents contextes du marché. Pour le contexte du bois d’œuvre,la gestion de la demande est d’autant plus complexe vu la structure divergente des proces-sus de production, l’instabilité de l’approvisionnement, l’incertitude et les fluctuations de lademande et des prix, ainsi que l’hétérogénéité des clients.

1.2.1 Cas d’étude

La Figure 1.1 illustre le cas industriel étudié dans cette thèse. On a considéré une entreprisedu bois d’œuvre composée de trois scieries de même capacité localisées dans la province deQuébec et approvisionnées en matières premières par deux sources d’approvisionnement.Ces trois scieries sont dotées de ressources de sciage, de séchage et de rabotage permettantde générer des produits verts et secs de caractéristiques variées (dimensions, longueurs etgrades). Inspiré du cas de la région du Québec, l’entreprise cible des clients localisés dans dif-férents marchés, soit principalement les États-Unis comme le principal marché et des régionscanadiennes comme un marché local. D’autres marchés internationaux, comme la Chine etle Japon, peuvent être considérés.

FIGURE 1.1 – Réseau d’approvisionnement d’une entreprise du bois d’œuvre

Les clients provenant de ces différents marchés peuvent être classifiés selon différentes carac-téristiques (Gaston and Robichaud 2017). La Figure 1.2 illustre une classification des clientsdu bois d’œuvre selon leurs sensibilités aux prix et à la qualité.

3

FIGURE 1.2 – Différents types de clients du bois d’œuvre

Contrairement aux processus de production usuels, l’industrie du bois d’œuvre se caracté-rise par des processus de production complexes. En effet, pour un produit en entrée, plu-sieurs types de produits sont fabriqués (flux divergents), et cela de manière simultanée (co-production). De plus, pour un produit en entrée, plusieurs recettes et plans de coupes (voirpar exemple la Figure 1.3) sont possibles, ce qui influence les proportions de chaque produitqui sera obtenu (Gaudreault et al. 2010, Rafiei et al. 2014).

FIGURE 1.3 – Exemple de plans de coupe. Adapté de Vila, Martel, and Beauregard (2006)

4

Dans cette thèse, on considère différents produits de commodité. Contrairement aux pro-duits personnalisés, les produits de commodité se caractérisent par une forte demande, quidépasse dans la plupart des cas ce qu’une seule scierie peut offrir, ceci compte tenu sescontraintes de production et d’approvisionnement. De plus, les prix et la demande des pro-duits de commodité du bois d’œuvre sont connus par leurs fluctuations cycliques reliées àla saisonnalité des activités de construction.

1.2.2 Questions de recherche

Tout au long de cette thèse, nous nous intéressons à répondre à la question de recherche sui-vante : Comment un processus multiniveau de gestion de la demande peut être efficacementmanagé dans un contexte de capacité limitée, tel est le cas des entreprises du bois d’œuvre,afin de maximiser les revenus et de garantir des niveaux du service élevés aux clients prio-ritaires?

Afin d’aborder cette problématique générale, nous proposons de répondre aux questions derecherche suivantes liées aux approches intégrées de gestion de la demande dans un contextede capacité limitée :

Question 1 : Comment peut-on formuler le problème de gestion de la demande comme unprocessus intégré considérant les décisions tactiques et opérationnels, ainsi que les décisionsde ventes à temps réel ? Quelles sont les différentes interactions possibles entre ces troisniveaux de décision?

Question 2 : Quelles approches de gestion de la demande peut-on intégrer afin de maximiserles revenus dans un contexte de capacité limitée et de profiter des opportunités du marché(soit dans le cas de l’industrie du bois d’œuvre : la saisonnalité de la demande et des prix,ainsi que l’hétérogénéité des clients) ?

Question 3 : Comment les situations du marché du bois d’œuvre peuvent-elles affecter laperformance du processus de gestion de la demande? Quels sont les facteurs les plus perti-nents?

Question 4 : Peut-on améliorer la performance du processus intégré de gestion de la de-mande en incorporant des pratiques communes dans l’industrie du bois d’œuvre, telles quela substitution des produits et la satisfaction partielle des commandes ("partial fulfillment") ?

Ces questions ont été peu abordées dans la littérature et l’objectif de cette thèse est d’appor-ter des réponses à chacune d’entre elles à travers trois contributions de recherche, tel queprésenté dans la Figure 1.4, ceci en considérant le cas de l’industrie du bois d’œuvre. Dansce qui suit, on définira quelques concepts clés, puis on détaillera chacune des contributionsde la thèse.

5

FIGURE 1.4 – Structure et contributions de la thèse

6

1.3 Concepts préliminaires

Dans cette section, nous présenterons les différents concepts abordés dans cette thèse et lesprincipaux travaux de recherche étroitement liés avec les trois contributions présentées.

1.3.1 Gestion de la demande

La gestion de la demande inclue toutes les activités permettant de créer un flux de demandecoordonné (ou synchronisé) entre la chaîne d’approvisionnement et ses différents marchés(Stadtler and Kilger 2005). Comme le montre la Figure 1.5, ce processus va au-delà de la fonc-tion traditionnelle du marketing, soit la création et la simulation de la demande et prend encharges les tâches suivantes : anticiper la demande via les outils de prévision, assurer la coor-dination entre les différentes fonctions internes de l’entreprise (exemple : production, distri-bution . . . ) et la collaboration des différentes entreprises de la chaîne d’approvisionnement,évaluer la contribution en termes de profit des différentes demandes (différents produits /clients / périodes) et gérer efficacement l’allocation de la capacité en mettant l’accent sur lesdemandes les plus profitables(Crum and Palmatier 2003, Mentzer, Myers, and Stank 2007).

FIGURE 1.5 – La position de la gestion de la demande dans la gestion de la chaîned’approvisionnement. Adapté de Mentzer, Myers, and Stank (2007)

Évolution paradigmatique de la gestion de la demande

La gestion de la demande se limitait au début au fait de prévoir la demande (Mentzer, Myers,and Stank 2007). Cette définition a évolué au cours du temps sur deux volets : le premier vo-let concerne la progression des techniques influençant la demande telles que les stratégies demarketing et d’établissement de prix ("pricing"). Parallèlement à ce premier volet s’ajoute un

7

deuxième volet, soit la communication des informations sur la demande (Kaipia, Korhonen,and Hartiala 2006). Cette coordination a débuté au niveau de la chaîne interne de l’organi-sation. A partir des années 90, avec l’apparition des concepts de chaîne / réseau d’approvi-sionnement et l’évolution des outils d’aide à la décision et des technologies d’information(le passage des MRPI / MRP II dans les années 80 vers les systèmes de planification avancésASP et les progiciels de gestion intégrés ERP dans les années 90 (McGaughey and Gunaseka-ran 2007)), il est de plus en plus possible de coordonner et de collaborer avec les différentesparties prenantes de la chaîne / réseau d’approvisionnement via des processus de coordi-nation tels que la planification des ventes et des opérations (Sales and Operations PlanningS&OP).

Le besoin d’efficience et de collaboration en termes de visibilité de la demande en aval dela chaîne d’approvisionnement (comme avoir accès aux données des points de ventes et auxdonnées d’inventaires tout le long de la chaîne d’approvisionnement) a donné naissance àun autre concept, soit la gestion de la chaîne de la demande ou " demand chain manage-ment" (Frohlich and Westbrook 2002, Wu, J. Gao, and R. Yu 2009). Vu le besoin croissantde personnalisation, l’objectif est de comprendre, influencer et gérer la demande, ainsi qued’assurer l’agilité tout le long de la chaîne d’approvisionnement (Duarte Canever, Van Trijp,and Beers 2008). Il s’agit i) d’acquérir des informations plus détaillées et plus fiables sur lesconsommateurs et de les transmettre aux partenaires le long de la chaîne de la demande afind’améliorer la précision des prévisions de la demande ("collaborative forecasting" ), ii) demettre l’accent sur le client et se baser sur la compréhension de la demande pour établir lesstratégies et les plans de toute la chaîne d’approvisionnement et iii) d’atteindre les cibles desventes tout en utilisant efficacement la capacité et les ressources.

Prévisions de la demande vs. Plans et cibles de ventes

On définit les prévisions de la demande comme étant la projection dans le futur de la de-mande qu’on peut recevoir, considérant un ensemble de conditions du marché. Ces prévi-sions doivent être distinguées des plans des ventes qu’on prévoit concrétiser sous réserve dela réalisation de l’ensemble des plans de production, approvisionnement, distribution, etc.On peut également parler de cibles de ventes, qui représentent les niveaux de ventes établiescomme motivation pour l’équipe des ventes et de marketing. Le rôle de l’ensemble des plansest de gérer efficacement la capacité de la chaîne d’approvisionnement, en se basant sur lesprévisions de la demande, afin d’atteindre ou dépasser les cibles de ventes (Mentzer, Myers,and Stank 2007).

8

1.3.2 Planification des ventes et des opérations (Sales and Operations PlanningS&OP)

La planification des ventes et des opérations (Sales and Operations Planning S&OP) permetd’intégrer le plan de production et le plan de ventes établi en se basant sur les prévisions dela demande (Olhager, Rudberg, and Wikner 2001), tel qu’illustré par la Figure 1.6. Le S&OPpeut être défini comme un processus périodique de planification tactique qui lie verticale-ment les plans d’affaires et les plans stratégiques à long terme avec les plans opérationnelsà court terme, et horizontalement la demande avec les capacités de la chaîne d’approvision-nement (Feng, D' Amours, and Beauregard 2008). Selon l’APICS (2013), le S&OP intègrel’ensemble des plans d’affaires d’une entreprise (approvisionnement, production, ventes,clients, marketing, R&D et finances) dans un plan général, facilite la coordination entre lesdifférentes fonctions et supporte les plans stratégiques et les plans d’affaires en couvrantun horizon de planification entre un an et deux ans. Affonso, Marcotte, and Grabot (2008)affirment que la longueur de l’horizon de planification doit correspondre au moins à la lon-gueur de l’horizon budgétaire.

FIGURE 1.6 – La coordination via la planification des ventes et des opérations. Adapté deMentzer, Myers, and Stank (2007)

Aspect organisationnel du S&OP

Le S&OP assure l’opérationnalisation des décisions stratégiques, comme par exemple les dé-cisions de capacité. Dans ce contexte, Olhager, Rudberg, and Wikner (2001) ont traité un casoù l’espace de décision du S&OP est affecté par une stratégie d’expansion ou de réduction dela capacité. Ils ont exposé comment le processus S&OP peut influencer la planification de la

9

capacité à long terme par la réévaluation de l’exécution des plans de ventes et de production,en identifiant périodiquement le besoin d’augmenter ou diminuer le niveau de capacité.

Un processus S&OP peut aussi agir comme un mécanisme continu qui supporte l’intégrationinter-fonctionnelle. Une étude de cas d’Oliva and Watson (2011) a montré que la fiabilité desinformations, la qualité des procédures de planification et la qualité de la coordination entreles différentes fonctions de l’entreprise représentent les principaux attributs qui supportentles fonctionnalités du processus et qui explorent le rôle de l’engagement constructif entre lesfonctions de l’entreprise. L’étude a montré qu’après trois ans de l’implantation d’un proces-sus S&OP dans une entreprise d’électronique, un bon niveau de performance a été atteint dûà l’amélioration des prévisions et à l’efficacité opérationnelle. Le processus S&OP a facilitéla planification intégrée de la chaîne d’approvisionnement et a poussé toutes les fonctions às’engager activement dans chaque étape du processus grâce à un aspect critique constant.

Le S&OP supporte également l’intégration entre les entreprises dans la chaîne d’approvi-sionnement. Dans ce sens, Affonso, Marcotte, and Grabot (2008) ont proposé un modèle deS&OP composé de trois niveaux liant le département commercial d’une entreprise avec lesservices d’achat de ses clients et le département d’achat avec les départements commerciauxde ses fournisseurs. Ils ont considéré les délais entre ces trois niveaux comme éléments clés.

L’efficacité du processus S&OP est soutenue par l’intégration des flux financiers et des fluxphysiques. Z. Wang and Hsu (2010) ont présenté un modèle du processus S&OP qui intègreles termes de payement. La simulation du modèle a montré que le délai de paiement repré-sente le principal facteur affectant la situation de trésorerie et que le processus S&OP estd’autant plus efficace qu’on diminue ce délai.

Gestion de l’incertitude

Dans un contexte incertain, le S&OP permet d’aligner les cibles de vente avec la disponibi-lité des ressources. En particulier, Chen-Ritzo et al. (2010) se sont intéressés à traiter expli-citement l’incertitude de la configuration des produits commandés dans un environnement"Configurer sur commande" (Configure-To-Order). Leur modèle de S&OP permet de sup-porter les décisions associées aux étapes de planification de l’approvisionnement et de révi-sion de la demande et de l’approvisionnement selon la flexibilité des fournisseurs. Ces deuxétapes ont été formulées comme deux programmes stochastiques avec recours, puis résoluesavec l’approche d’approximation moyenne de l’échantillon (SAA).

Le S&OP joue un rôle important de médiation pour améliorer la performance opérationnelledans les environnements de production caractérisés par une incertitude du marché (Olhagerand Selldin 2007). Dans ce contexte, Sodhi and Tang (2011) ont proposé un modèle de pro-grammation linéaire stochastique qui minimise les indicateurs de risque associés aux risquesd’arrérages, d’excès de stocks et d’emprunts excessifs. Plus récemment, Feng et al. (2013) ont

10

traité un problème de S&OP stochastique considérant les changements de tendance de l’éco-nomie, du marché et de l’approvisionnement dans l’industrie forestière. Ces changementsont été traduits par un ensemble de scénarios possibles. Les auteurs ont proposé un modèlede programmation stochastique à deux étages avec recours fixes afin de maximiser le pro-fit du producteur. Le premier étage lui permet de choisir, d’une part, les contrats de clientsqu’il peut satisfaire, et d’autre part les contrats d’approvisionnement qui lui garantissentses besoins en matière première. Le deuxième étage concerne les décisions opérationnellesà prendre compte tenu des décisions liées aux contrats signés. Une étude par simulation amontré que, comparé au modèle déterministe, le modèle stochastique est plus approprié etplus robuste face aux problèmes de décisions de contrats.

En outre, le processus S&OP permet d’atténuer les effets des erreurs de prévision. Grâce àune étude de cas dans l’industrie des panneaux à lamelles orientées, Feng, D' Amours, andBeauregard (2008) ont modélisé trois modèles de gestion de la chaîne d’approvisionnement :un premier modèle multisite basé sur un processus S&OP, un deuxième modèle où seulsle plan de ventes et le plan de production sont intégrés et un troisième modèle où toutesles fonctions de la chaîne d’approvisionnement sont planifiées séparément. La simulationde ces modèles face à une demande déterministe (Feng, D' Amours, and Beauregard 2008),puis face une demande stochastique en considérant une planification sur un horizon rou-lant (Feng, D' Amours, and Beauregard 2010), a démontré qu’on peut réaliser une meilleureperformance financière avec le modèle basé sur un processus S&OP.

Défis du S&OP

Plusieurs enquêtes (par exemple IBFP and APICS (2011)) et articles de recherche (voir lesrevues systématiques of Thomé et al. (2012) and Tuomikangas and Kaipia (2014) pour uneliste exhaustive) présentent le S&OP comme un processus exigeant un changement radicalde la culture de l’organisation. Certes, l’implantation du S&OP fait face à plusieurs défis. Pre-mièrement, il faut noter que les objectifs des différentes fonctions de l’organisation peuventêtre contradictoires puisque les décisions optimales d’une planification centralisée peuventdifférer des décisions optimales obtenues si on planifie chaque fonction séparément (Feng,D' Amours, and Beauregard 2008). Deuxièmement, il n’est pas toujours évident d’atteindrele niveau d’intégration souhaité et de faire un compromis entre différents choix tranchants(Oliva and Watson 2011, Sodhi and Tang 2011), comme avoir à considérer le risque de sur-plus de production vs. le risque d’être en pénurie. A cela s’ajoute la difficulté de gérer lesincertitudes (Feng, D' Amours, and Beauregard 2010, Chen-Ritzo et al. 2010) et l’intégrationinter-entreprises (Affonso, Marcotte, and Grabot 2008), ainsi que la complexité de mesurerla performance d’un tel processus (Hulthén, Naslund, and Norrman 2017). Dans ce sens,Thomé et al. (2012) and Tuomikangas and Kaipia (2014) témoignent de l’importance de larecherche académique afin de surmonter ces défis, et particulièrement les études de cas dans

11

différents contextes à l’instar de Oliva and Watson (2011) dans l’industrie électronique, Lim,Alpan, and Penz (2017) dans l’industrie automobile, Wery et al. (2018) dans l’industrie fores-tière, etc.

Les revues de Thomé et al. (2012), Tuomikangas and Kaipia (2014) et Kristensen and Jonsson(2018) proposent de considérer dans les recherches futures sur le S&OP la complexité et lescaractéristiques spécifiques à chaque industrie. En particulier, cette thèse met l’accent, d’unepart, sur la valeur d’implanter un tel processus dans l’industrie du bois d’œuvre caractériséepar des processus divergents, une demande saisonnière et des clients hétérogènes. D’autrepart, elle souligne l’intérêt d’intégrer le S&OP avec les concepts de gestion des revenus, cequi n’a pas été abordé auparavant dans la la littérature.

1.3.3 Satisfaction de la demande, promesse de livraison et gestion des revenus

Satisfaction de la demande et promesse de livraison

Contrairement au processus S&OP qui concilie sur le moyen terme le plan de ventes avecles capacités de l’entreprise (voir la Figure 1.7), la satisfaction de la demande ("Demand ful-fillment") permet de traiter les commandes en temps réel et de générer des promesses delivraison ("Order promising") en se basant sur les informations sur la capacité (Stadtler andKilger 2005).

FIGURE 1.7 – Différents niveaux de planification. Adapté de Feng, D' Amours, andBeauregard (2008)

12

À partir des prévisions de la demande et en tenant compte des contraintes de la capacité,il est possible de calculer les quantités disponibles à la vente (Available-To-Promise ATP) àun moment donné dans le temps, ce qui constitue la portion d’inventaire et de productionplanifiée maintenue par le calendrier directeur de production et des plans des besoins ma-tières. Connaissant l’ATP, des décisions d’allocation aux commandes doivent être prises. Cesdécisions dépendent de la position du point de découplage (Fleischmann and Meyr 2003),séparant les parties de la chaîne d’approvisionnement pilotées par prévision des parties pi-lotées par les commandes (voir la Figure 1.8).

FIGURE 1.8 – Point de découplage selon le type d’environnement de production. Adapté deFleischmann and Meyr (2003)

Dans un environnement de fabrication pour les stocks (Make-To-stock MTS) où les processusd’approvisionnement et de production sont pilotés par prévision, l’allocation d’ATP se faitau niveau des produits finis selon différentes règles d’allocation. Dans la littérature, on faitsouvent référence à l’étude de Pibernik (2006). Ce dernier a fourni des heuristiques et desprogrammes linéaires associés à des logiques d’allocation simples et a analysé le potentielde ces mécanismes à contribuer à une gestion efficace de l’allocation de l’ATP dans unesituation de rupture de stock.

Gestion des revenus (Revenue management RM) pour les systèmes manufacturiers

Dans un mode de contrainte de l’approvisionnement où toute la demande ne peut pas êtrecomblée, plusieurs recherches se sont basées sur les concepts de RM afin de résoudre le pro-blème d’allocation d’ATP. Le RM est considéré comme une méthode optimale permettantd’assurer des prix moyens de vente plus élevés et une relation plus forte avec les clientsayant une volonté à payer (Stadtler and Kilger 2005). Selon Phillips (2005), le RM est définicomme l’ensemble de stratégies et de techniques permettant de gérer l’allocation des capa-

13

cités aux différentes classes de clients et de conserver les capacités réservées à chaque classeen discernant les limites de réservations de chaque classe dans l’objectif de maximiser lesrevenus.

Le RM a connu des grands succès dans les entreprises de service, notamment dans les com-pagnies aériennes, les hôtels, etc. Quante, Meyr, and Fleischmann (2009) ont été les premiersà explorer les domaines d’application du RM dans un environnement manufacturier. Danscet article, ils ont différencié entre l’approche traditionnelle de RM où l’approvisionnementest non considéré et l’approche par allocation du disponible à la vente (allocated Available-To-Promise aATP) où l’approvisionnement est une donnée exogène. Inspiré du RM, le prin-cipe d’aATP consiste à allouer les ressources rares de l’entreprise aux classes de clients priori-taires dans le but de maximiser les revenus. La planification d’aATP dans le cas déterministeest généralement composée de deux étapes : la planification de l’allocation et la consomma-tion en temps réel. La première étape se fait à moyen terme et consiste à affecter l’ATP à unnombre défini de classes de clients. Les quantités allouées aux différentes classes sont par lasuite consommées à temps réel par la deuxième étape (Meyr 2009, Azevedo, D' Amours, andRönnqvist 2016).

Dans la littérature, on a procédé différemment pour identifier les classes de clients. Meyr(2009) a utilisé une mesure artificielle qui décrit l’importance de l’ordre, remplaçant ainsiles règles d’allocation basées sur la maximisation de profit et permettant de différencier lesclients et de répondre aux commandes de façon immédiate. Il a appliqué cette approchedans un environnement MTS, en considérant un seul produit dans un site de l’industrie duluminaire.

En partant de cette étude, Azevedo, D' Amours, and Rönnqvist (2016) ont appliqué l’ap-proche d’aATP pour une entreprise du bois d’œuvre à l’Est du Canada, distribuant ses pro-duits principalement à l’Est du Canada et aux marchés Nord-américains. La clientèle a étéclassifiée en six segments. En effet, l’Est canadien et la zone des Grands Lacs d’Amériquedu Nord ont été considérés comme deux divisions géographiques différentes à cause de ladifférence des comportements de prix et des coûts de transport et de la fluctuation du tauxde change entre le dollar américain et le dollar canadien. Chaque division a été aussi décom-posée en trois niveaux de sensibilité des clients au prix.

En considérant plusieurs produits, Azevedo, D' Amours, and Rönnqvist (2016) ont déve-loppé un modèle de promesse de livraison basé sur l’aATP afin d’optimiser le revenu nettotal de la compagnie, représentée comme un environnement MTS multisite. Dans une pre-mière étape, la planification de l’allocation permet de réserver de l’ATP aux différentesclasses de clients en se basant sur des informations déterministes liées aux prévisions dela demande et du prix. L’horizon de planification de l’allocation considéré est de quatrepériodes. Dans une deuxième étape, les allocations sont consommées à temps réel par les

14

commandes arrivant de façon aléatoire. Pour chaque commande reçue, l’approvisionnementd’ATP se fait à partir du nœud le plus proche, de façon à ce que tous les produits d’unecommande soient retirés du même nœud. Dans le cas où la demande dépasse à la quan-tité allouée dans la première étape, la priorité est accordée aux segments de clients les plusrentables.

En réalité, la demande n’est pas déterministe. Par conséquent, Quante, Fleischmann, andMeyr (2009) ont traité le cas stochastique pour plusieurs classes hétérogènes de clients et ontdéveloppé un modèle qui permet de décider si l’ordre est accepté, rejeté ou mis en attenteavec une pénalité de retard. L’étude a permis de caractériser une politique de satisfactionde la demande optimale, puis de la comparer avec la politique d’allocation déterministedéveloppée par Meyr (2009). Contrairement à l’étude de Quante, Fleischmann, and Meyr(2009) qui a supposé que chaque commande doit être livrée à la même période où elle a étéreçue, on a considéré dans cette thèse des délais de livraison qui suivent une loi triangulairede paramètres variables selon le type de client.

Dans les environnements ATO (Assemblate-To-Order ou assembler sur commande) et CTO(Configure-To-Order ou configurer sur commande), la production des composants est pilo-tée par prévision. L’allocation d’ATP est donc appliquée au niveau des composants. Tsai andWang (2009) sont parmi les premiers à considérer les concepts de gestion des revenus dansun contexte similaire. Récemment, Guhlich, Fleischmann, and Stolletz (2015) ont développéune heuristique d’enchères pour un système de production ATO faisant face à une demandestochastique.

Dans les environnements MTO (Make-To-Order ou produire sur commande), l’approvision-nement se fait à partir des prévisions, tandis que la production, l’assemblage et la distribu-tion dépendent de la réception des commandes. L’ATP dépend ainsi de la disponibilité descomposants et des contraintes de capacité afin de permettre de générer une date promise.Dans ce contexte, Spengler, Rehkopf, and Volling (2007) ont développé une approche de RMpour améliorer la sélection des commandes à court terme dans l’industrie de fer et acier. Desétudes plus récentes ont utilisé la programmation dynamique et des modèles de RM baséssur des politiques de fixation de prix ("pricing") dans des environnements ATO et MTO. Vol-ling et al. (2011), par exemple, ont traité le cas d’un seul site, contrairement à Tsai and Wang(2009) qui ont considéré le cas multisite.

Dans toutes les études présentées, l’application de la gestion des revenus est abordée enconsidérant un plan fixe d’approvisionnement/production et un horizon de planification àcourt terme, ignorant ainsi le profit potentiel qui peut être généré en anticipant la demande àmoyen terme. Cette thèse permet, grâce au S&OP, d’inclure les décisions tactiques d’appro-visionnement et de production afin de profiter de la saisonnalité des prix et de la demande.Ces décisions sont réajustées en considérant un horizon roulant de planification.

15

1.3.4 Substitution des produits

La substitution des produits est une pratique commune qui vise à faire face aux pénuriesdans un contexte de capacité limitée et à faire écouler le stock (particulièrement, dans le casde coproduction). Cette pratique présente un potentiel d’efficacité (Lang 2010) puisqu’ellepermet d’assurer un meilleur niveau de service (combler les pénuries en cas de rupture destock) et de réduire les coûts de stockage, les coûts de mise en route (produire en grandslots des produits pouvant être des substituants pour d’autres produits), les coûts unitaires(produire des substituants dont le coût est inférieur que le produit original) et les gaspillages(par exemple, dans le cas des produits périssables).

Parmi la littérature abondante sur les modèles de substitution, Lang (2010) a présenté unevue d’ensemble sur les modèles de gestion de stocks et de production avec substitution.En particulier, il a illustré comment la substitution des produits peut être introduite dansle cadre décisionnel de la chaîne d’approvisionnement. De plus, il a discuté des options desubstitution dans différents problèmes de lotissement et d’ordonnancement.

La décision de substitution peut être prise au niveau tactique. Ervolina et al. (2009), parexemple, ont proposé un modèle tactique d’allocation dans un contexte ATO. Ce modèlepermet de considérer le disponible à la vente (ATP) d’un produit donné, ainsi que l’ATP detous ses substituants. L’étude de cas dans une entreprise de production d’ordinateurs, offrantplusieurs alternatives de substitution, a démontré un grand potentiel à gérer les surpluset à réduire les couts de stockage. Cette étude a considéré les plans de production commedonnées exogènes, contrairement aux travaux de cette thèse.

Le concept de substitution au niveau opérationnel a été largement abordé dans la littéra-ture de la gestion des revenus (RM), mais plus pour le secteur de service (voir par exemple,Talluri and Van Ryzin (2004), Phillips (2005), Petrick et al. (2010, 2012), Gönsch, Koch, andSteinhardt (2014)). En particulier, Steinhardt and Gönsch (2012) ont proposé différentes for-mulations dynamiques de contrôle de la capacité en considérant l’option de sur-classement("upgrading"). Les auteurs ont démontré qu’intégrer le RM et le sur-classement performemieux que les approches de RM conventionnelles.

Dans un contexte manufacturier, la plupart des études ont traité la substitution comme uncomportement des clients en conséquence des décisions des prix ("pricing") : Gurler, Oztop,and Sen (2009) ont traité le cas d’un détaillant offrant deux produits périssables et substi-tuables pour une seule période. Ils ont considéré la demande comme fonction des prix desdeux produits et ont démontré que le profit optimal dépend de la relation de corrélationentre la demande des deux produits. Sibdari and Pyke (2010) se sont intéressés au cas dedeux entreprises concurrentes offrant des produits substituables. Ils ont proposé un modèledynamique utilisant la théorie de jeu pour décider des prix des produits pour chaque en-treprise. Kim and Bell (2011), quant à eux, ont investigué l’impact de la substitution créée

16

par la différence des prix des produits substituables sur les décisions optimales de prix et deproduction pour le cas d’une seule période. Contrairement à ces trois études, notre troisièmecontribution considère la substitution comme une décision de l’entreprise et investigue l’im-pact de cette pratique dans un environnement multipériode grâce à la planification en hori-zon roulant. Dans notre cas, nous analysons la possibilité de vendre au client un produit dequalité supérieure à la qualité du produit demandé.

1.3.5 Évaluation des systèmes industriels

Simulation en horizon roulant

La planification des systèmes industriels est un problème complexe incorporant des dé-cisions interreliées à différents niveaux (Kibira, Shao, and Johansson 2016). Plusieurs mé-thodes utilisées pour évaluer les processus de planification font l’hypothèse d’un environne-ment statique, alors les plans développés peuvent ne pas être optimaux dans un contexte oùdifférents événements se manifestent de façon dynamique. Dans ce contexte, la simulationen horizon roulant est largement utilisée afin de capter l’évolution des données à temps réelau cours du temps.

Dans la littérature, plusieurs études sur des problèmes de lotissement et d’ordonnance-ment (voir par exemple Meixell (2005), Al-Ameri, Shah, and Papageorgiou (2008), Torkaman,Ghomi, and Karimi (2017)) ont démontré que la planification en horizon roulant permetde générer des solutions plus réalistes que la planification à horizon fixe, et en particulierdans un contexte d’incertitude (voir par exemple Boulaksil (2016), Knoblich, Heavey, andWilliams (2015), Quddus et al. (2017), Rahdar, Wang, and Hu (2018)). Dans le contexte fores-tier, Feng, D' Amours, and Beauregard (2010) ont présenté des modèles de simulation en ho-rizon roulant afin d’analyser de la performance d’un processus de planification des ventes etdes opérations (S&OP) entièrement ou partiellement intégrée. L’étude a également comparéla performance des modèles en horizon roulant à la performance des modèles déterministesà horizon fixe et a montré que, malgré l’importance des modèles déterministes pour la re-cherche théorique, ils ne sont pas suffisants pour le support à la décision et l’évaluation desperformances dans un environnement d’affaires réel. Une étude plus récente de Rafiei et al.(2014) a proposé une plateforme d’optimisation et de simulation en horizon roulant permet-tant de comparer différentes politiques de re-planification périodique de la production, touten tenant compte des caractéristiques complexes de l’industrie de la seconde transformationdu bois. Dans cette thèse, on a utilisé la simulation en horizon roulant afin de prendre encompte les engagements de vente et les niveaux d’inventaires lors de la re-planification.

Planification d’expériences et métamodèles

Afin de conduire la simulation de manière efficace et de générer des conclusions avec leminimum d’essais possible, on a besoin d’une démarche rigoureuse de réflexion et d’analyse.

17

La planification d’expériences est l’art d’organisation des essais et de leur enchaînement aucours de l’expérimentation (Goupy 2015). Cet outil a été largement utilisé dans différentsdomaines de recherche comme la chimie, la gestion de qualité, etc (Kleijnen et al. 2005).Dans le contexte de gestion de la chaine d’approvisionnement, les plans factoriels et les plansTaguchi ont été les plus utilisées (voir par exemple, Bottani and Montanari (2010), Sandhu,Helo, and Kristianto (2013), Dev, Shankar, and Debnath (2014) et Hussain, Khan, and Sabir(2016) pour des problèmes de gestion de stock, Nedaei and Mahlooji (2014) et Assarzadeganand Rasti-Barzoki (2016) pour des problèmes d’ordonnancement et Santa-Eulalia et al. (2011)et Olaitan and Geraghty (2013) pour le contrôle de la production).

Pour l’évaluation des systèmes de simulation dans un contexte de gestion de la chaîne d’ap-provisionnement, plusieurs chercheurs, notamment Kleijnen et al. (2005) et Law (2015), re-commandent fortement d’utiliser les plans de remplissage d’espace ("space-filling designs")pour deux principales raisons. Premièrement, ce type de plans est flexible pour la concep-tion puisqu’il impose peu de restrictions sur les facteurs et les valeurs, tout en minimisant lenombre d’expériences. Deuxièmement, le concept de remplissage d’espace facilite l’analyseen permettant de construire différents types de métamodèles complexes et de considérer lesinteractions à des degrés élevés vu que pour chaque couple de facteurs, on trouve diffé-rentes combinaisons de valeurs. La Figure 1.9 montre un exemple de cas de deux facteurs etpermet de comparer l’espace couvert (càd, les différentes combinaisons de valeurs des deuxfacteurs) par un plan de remplissage d’espace comparé à un plan factoriel.

Afin d’analyser les résultats générés par un plan d’expériences, l’usage de métamodèles esttrès répandu dans la littérature en simulation. En effet, un métamodèle permet de prévoirles résultats d’un modèle de simulation en réponse à un ensemble de paramètres donné.Il permet ainsi de projeter le comportement du modèle de simulation face aux différentesconfigurations possibles (Law 2015). La technique de krigeage ("kriging") est une approched’interpolation recommandée pour les modèles aléatoires tels que les modèles de simulation(Law 2015). Elle permet de générer un métamodèle faisant une approximation globale de lafonction entrées/sorties (contrairement aux métamodèles polynomiaux où l’approximationest locale), ce qui est adapté aux modèles de simulation qui couvrent un espace assez large(Kleijnen 2017).

Dans cette thèse, on est amené à faire des expériences coûteuses en terme de temps. Spé-cifiquement dans notre deuxième contribution, plusieurs facteurs ont été considérés. On adonc utilisé un plan de remplissage d’espace et des métamodèles de krigeage afin d’obtenirle maximum d’informations tout en minimisant le nombre d’expériences. Dans la littérature,cette procédure a été peu utilisée afin d’évaluer un système industriel complexe tel que celuiétudié dans cette thèse.

18

FIGURE 1.9 – Comparaison d’un plan factoriel à un plan de remplissage d’espace pour lecas de deux facteurs. Adapté de Soderborg (2009)

1.4 Contributions de recherche et structure de la thèse

Afin de répondre aux différentes problématiques évoquées dans la Section 1.2, nous propo-sons trois contributions. Le but est d’amener de nouvelles approches intégrées de gestion dela demande pour supporter la prise de décision dans les entreprises du bois d’œuvre. Cescontributions ont été structurées dans ce rapport de thèse sous forme de trois chapitres.

1.4.1 Chapitre 2

Le chapitre 2 présente le premier article de cette thèse, intitulé "Integrating Revenue Ma-nagement and Sales and Operations Planning in a Make-To-Stock environment : Softwoodlumber case study". Premièrement, un cadre décisionnel permettant de supporter les respon-sables des ventes dans un environnement MTS est fourni (voir la Figure 1.10). Un proces-sus pour la gestion de la demande faisant intervenir les différentes fonctions de l’entrepriseest proposé. Il permet de mettre en évidence les rétroactions entre les niveaux de planifica-tion, ainsi que les entrées/sorties de chaque étape du processus. Ce processus intègre deuxconcepts communs en gestion de la demande (à notre connaissance, aucune publication n’aétudié la valeur de l’intégration de ces deux concepts ensemble), soit la planification desventes et des opérations (S&OP) et la gestion des revenus (RM). Sur un horizon roulantd’une année, on considère les décisions à prendre sur différents niveaux de planification et

19

FIGURE 1.10 – Éléments de recherche du premier article

à différentes fréquences dans le cas d’une entreprise composée de plusieurs sites de produc-tion et visant des segments hétérogènes de clients.

Deuxièmement, on reformule deux modèles de programmation linéaire (un modèle de S&OPpour les entreprises du bois d’œuvre et un modèle de promesse de livraison basée sur lesconcepts RM) déjà fournis par des études antérieures afin d’introduire la planification desallocations et de permettre la réaffectation des commandes. En effet, l’objectif est d’offrir lapossibilité de changer en temps réel les décisions d’allocation, la manière avec laquelle lescommandes seront satisfaites et à partir de quels sites elles seront livrées tant que les com-mandes fermes ne sont pas expédiées. Une plateforme d’optimisation et de simulation estdéveloppée afin d’assurer la planification en horizon roulant.

Finalement, on évalue les bénéfices d’intégrer le S&OP et le RM en considérant comme in-dicateurs le profit annuel, les ventes annuelles et le pourcentage de satisfaction des clientsprioritaires (les clients les plus payants dans notre cas d’étude). On propose de réaliser unesimulation en horizon roulant sur une année et de comparer la performance du processusintégré aux performances des processus conventionnels de gestion de la demande consi-dérant le S&OP et le RM de façon disjointe (soit à un processus considérant seulement le

20

S&OP, un processus considérant seulement le RM et un processus considérant une simpleapproche premier-arrivé premier-servi pour la promesse de livraison sans aucune planifica-tion tactique). Les résultats de simulation sont valables sous réserve de certaines hypothèsesprésentées dans la Figure 1.10. La Figure synthétise également le problème, la méthodologieet les contributions du premier article.

1.4.2 Chapitre 3

Le chapitre 3 présente le deuxième article de cette thèse, intitulé "Configuration and evalua-tion of an integrated demand management process using a space-filling design and Krigingmetamodeling". Dans cette étude, on examine la capacité des processus intégrés de gestionde la demande à performer face à différentes situations du marché (voir la Figure 1.11).

FIGURE 1.11 – Éléments de recherche du deuxième article

L’idée est d’investiguer les effets de facteurs variés du marché (soit l’intensité de la demandecomparée à la capacité, la précision des prévisions, l’hétérogénéité des clients manifestée parla différence entre les prix offerts par les multiples segments de clients, la variabilité de lataille des commandes). A cette fin, on utilise la plateforme d’optimisation et de simulationdéveloppée dans le premier article afin d’évaluer, via une simulation en horizon roulant,la performance de différents processus intégrés de gestion de la demande. On propose de

21

considérer différentes séquences d’arrivée des commandes et de comparer deux approchesde promesse de livraison : une approche basée sur le concept premier-arrivé premier-serviet une deuxième approche basée sur les concepts de RM.

De point de vue méthodologique, on a suivi la procédure de Montgomery (2009) de designet d’analyse des expériences (une procédure adaptée de Coleman and Montgomery (1993)).On a commencé par identifier les objectifs des expériences et sélectionner les mesures deperformance qu’on va évaluer en se basant sur les objectifs des responsables des ventes dansles entreprises du bois d’œuvre. Ensuite, en partant de la littérature du S&OP et du RM dansun contexte manufacturier, on a sélectionné les facteurs les plus pertinents et on a défini lacatégorie / l’intervalle de valeurs que peut prendre chaque facteur. On a utilisé un plan deremplissage d’espace pour définir les différents scénarios du marché qu’on va simuler etdes métamodèles de krigeage pour analyser les résultats. Cette procédure est relativementnouvelle dans un contexte de gestion de la chaîne d’approvisionnement, ce qui constitue laprincipale contribution de l’article. Finalement, à partir de l’analyse des résultats de simu-lation, on a identifié les implications managériales et conclu par des recommandations pourles entreprises ouvrant dans un contexte de capacité limitée.

1.4.3 Chapitre 4

Dans ce chapitre, on présente un troisième article intitulé "Simulating an integrated reve-nue management approach for a coproduction system with product substitution". Tel queprésenté dans la Figure 1.12, cet article contribue à l’extension des applications du RM auxsystèmes manufacturiers. La substitution des produits étant une pratique commune dansl’industrie du bois d’œuvre, on vise dans cet article à analyser l’effet d’intégrer la substitu-tion et les concepts de RM. Cette intégration a été peu abordée dans des secteurs autres quele secteur de service.

Cette étude considère que le vendeur accepte de fournir un produit de meilleure qualité auprix du produit original demandé. Ceci est équivalent à un sur-classement ("upgrading")pour les entreprises de service. A cette fin, on formule un modèle de promesse de livrai-son générique qui peut être adapté pour différentes approches de promesse de livraison.Les modèles conventionnels assignent la demande aux allocations définies par le plan tac-tique, ces allocations étant définies par client et par période de livraison. Afin d’introduire lasubstitution des produits, on ajoute une troisième dimension, soit le produit.

En utilisant la plateforme d’optimisation et de simulation développée dans le premier article,on évalue, via la simulation en horizon roulant, les bénéfices d’intégrer les concepts de RMet la substitution des produits dans un contexte de capacité limitée et coproduction. Pour lesexpérimentations, on considère différents scénarios de demande et de prix.

22

FIGURE 1.12 – Éléments de recherche du troisième article

1.5 Conclusion

À travers ce premier chapitre, on a introduit le problème de gestion de la demande dansl’industrie du bois d’œuvre et le cas d’étude. Ensuite, on a présenté les concepts clés abordésdans cette thèse et les principaux travaux de recherche étroitement liés avec les trois contri-butions. Finalement, on a détaillé, pour chacune de ces contributions, la problématique, leshypothèses et la méthodologie utilisée.

La Figure 1.4 a synthétisé les différentes contributions de cette thèse présentées sous formed’articles dans les chapitres 2, 3 et 4. Dans la première contribution, un cadre décisionnelest proposé afin de supporter la prise des décisions des ventes sur différents niveaux deplanification de façon à maximiser les revenus et la satisfaction des clients prioritaires. Unepremière série d’expériences est réalisée afin d’évaluer les bénéfices d’intégrer le S&OP et leRM. La plateforme d’optimisation et de simulation développée à ce stage est utilisée pourla deuxième contribution et permet d’expérimenter la performance de différents processusintégrés de gestion de la demande face à une variété de scénarios du marché. En particulier,une procédure relativement nouvelle dans le contexte de gestion de la chaîne d’approvision-nement est proposée : on utilise un plan de remplissage d’espace pour définir les scénarios

23

à simuler et des métamodèles de krigeage pour analyser la relation entre les indicateurs deperformance et les différents facteurs étudiés. La troisième contribution, quant à elle, pro-pose d’intégrer une pratique commune dans l’industrie du bois d’œuvre, soit la substitutiondes produits, avec les concepts de gestion des revenus. Finalement, un dernier chapitre seradédié pour la conclusion générale et présentera quelques pistes de recherche pertinentes.

24

Chapitre 2

Intégration de la gestion des revenuset de la planification des ventes et desopérations dans un environnement defabrication pour les stocks : Cas del’industrie du bois d’œuvre

Cet article, intitulé "Integrating Revenue Management and Sales and Operations Planning in aMake-To-Stock environment : Softwood lumber case study", a pour auteurs Maha Ben Ali, SophieD’Amours, Jonathan Gaudreault et Marc-André Carle. Il a été soumis dans le journal "IN-FOR : Information Systems and Operational Research" en janvier 2018. La version présentéedans cette thèse est identique à la version soumise.

25

Résumé

L’application de la gestion des revenus dans un contexte manufacturier a été généralementabordée dans la littérature en considérant un plan d’approvisionnement/production fixe etun horizon de planification à court terme. Cette étude assure de plus une visibilité à moyenterme dans l’objectif de permettre aux entreprises à capacité limitée et faisant face à une de-mande saisonnière, de maximiser leurs profits et d’améliorer le niveau de satisfaction desclients les plus payants. A cette fin, on propose un processus de gestion de la demande in-tégrant la planification des ventes et des opérations avec la promesse de livraison basée surles concepts de gestion des revenus. En premier lieu, on définit un cadre décisionnel multi-niveau afin de supporter les décisions de ventes prises aux niveaux tactique et opérationnel,ainsi que les promesses de livraison conclues en temps réel. Le cas d’étude, inspiré des entre-prises du bois d’œuvre québecoises, considère différents clients hétérogènes, des processusde production divergents et plusieurs usines. En second lieu, on propose une formulationmathématique intégrant un modèle de planification des ventes et des opérations adapté àl’industrie du bois d’œuvre et un modèle de promesse de livraison utilisant des limites deréservation imbriquées. Cette nouvelle formulation rend possible la révision des promessesde livraison à chaque exécution, tout en respectant les anciens engagements de vente. Unesimulation en horizon roulant permet de mettre en évidence la valeur de l’intégration de laplanification des ventes et des opérations et de la gestion des revenus et d’évaluer la per-formance du processus intégré proposé dans des scénarios variés, en le comparant à desprocessus conventionnels de gestion de la demande.

26

Abstract

Most research regarding revenue management in manufacturing has considered only a short-term planning horizon, assuming supply and production data exogenously given. Moti-vated by the case of the Canadian softwood lumber industry, this paper offers additionallya medium-term visibility for firms with limited capacity and faced with seasonal markets.We propose a demand management process for Make-To-Stock environments, integratingsales and operations planning (S&OP) and order promising based on revenue managementconcepts. Given heterogeneous customers, divergent product structure and multiple sourc-ing locations in a multi-period context, we first define a multi-level decision framework inorder to support medium-term, short-term and real-time sales decisions in a way to max-imize profits and to enhance the service level offered to high-priority customers. We fur-ther propose a mathematical formulation integrating an S&OP network model in the Cana-dian softwood lumber industry and an order promising model using nested booking limits.This new formulation allows reviewing previous order promising decisions while respect-ing sales commitments. A rolling horizon simulation is used to evaluate the performance ofthe proposed process in various demand scenarios and provides evidence that better perfor-mances can be achieved compared to common demand management practices by integratingS&OP and revenue management concepts.

27

2.1 Introduction

Revenue management (RM) techniques have received notable attention in manufacturingas a powerful tool for order promising in supply-constrained environment. Although it is acritical task, order promising focuses on maximizing short-term revenue, often ignoring thepotential profit that can be obtained by anticipating mid-term demand. In contrast, sales andoperations planning (S&OP) focuses on mid-term revenue and offers the possibility of takingadvantage of demand and price fluctuations. Unfortunately, it seems that current studies andexisting systems that dealt separately with RM or S&OP, hardly capture the needs of salesmanagers. In fact, the integration between RM and S&OP is not well understood either intheory or practice.

In this research, the integration of RM and S&OP is motivated by the case of most Cana-dian softwood lumber firms, which fight ferociously to be more competitive when facing aset of business challenges : economic pressures, high operating costs, divergent processes,heterogeneous customers, limited raw material availability and capacity, and seasonal mar-ket. Based on multiple meetings with softwood lumber managers from the Eastern Canadianregion, we noted first the lack of synchronization between the different business units of soft-wood supply chain due to divergent product structure (i.e. from one log, it is not possible togenerate different products independently) and the highly heterogeneous nature of its rawmaterial, and second the ingenuous manner in which orders are fulfilled.The objectives ofthis paper are i) to offer guidance for such firms by extending the existing research in de-mand management for Make-To-Stock (MTS) manufacturing systems in a way to maximizeprofits and enhance the service level offered to high-priority customers, and ii) to provideevidence to managers of the value of integrating S&OP and RM.

More precisely, our contributions are as follows : First, in order to support sales decisionsthat have to be taken at multiple planning levels and at different frequencies (real-time,short-term and mid-term sales decisions), we define a demand management process integra-ting S&OP and RM and considering differentiated demand segments, divergent productionprocesses and multiple sourcing locations in a multi-period context. Second, we propose amathematical model integrating an S&OP network model in the Canadian softwood lumberindustry and an order promising model based on RM concepts. Our integrated model alsooffers the possibility of changing decisions of how confirmed orders have to be fulfilled aslate as possible, which we called order reassignment. Third, we evaluate the demand mana-gement process performance with various demand scenarios via a rolling horizon simula-tion. We emphasize the benefits of integrating S&OP and RM concepts for softwood lumbermanufacturers located in Eastern Canada, as compared to demand management commonpractices.

This paper is structured as follows : In Section 2.2, we restate the problem faced by Canadian

28

softwood lumber firms and define the paper’s positioning via an overview of recent studiesabout S&OP and RM applications in manufacturing context. Furthermore, in Section 2.3 wepropose a demand management process for MTS environments, including S&OP at the tacti-cal level and real-time order promising based on RM concepts at the operational/executionlevel . In Section 2.4, a tactical model and an order promising model are formulated as li-near programs (LP) so that order assignment may be changed as late as possible, althoughthe decision of accepting or refusing an order is instantaneous and definitive. A networkperspective is considered. Afterward, the illustrative case, based on softwood lumber ma-nufacturers located in Eastern Canada, and experiments that will be conducted are depictedin Section 2.5. In Section 2.6, we discuss the potential benefits of integrating S&OP and RMconcepts and the managerial implications. Finally, concluding remarks are provided in Sec-tion 2.7.

2.2 Problem statement and related literature

In this section, we first state the problem faced by sales managers in Canadian softwoodlumber firms. Then, we proceed with a literature review to describe the basis of S&OP andto analyze the current research state on RM in manufacturing context.

2.2.1 Demand management problem in Canadian softwood lumber industry

The softwood lumber industry is an important sector in the Canadian economy. It offersthousands of direct jobs and significant benefits supporting indirect jobs. This sector is alsoinvolved in the development of rural and remote communities in certain regions. Moreo-ver, softwood lumber accounts for 20% of the value of Canadian forest product exports 1,destined for domestic and international markets.

During recent years, this industry has faced various trade and economic pressures (Du-four 2007), including Canada—US softwood lumber agreements and fluctuations in theCanada—US exchange rate, American anti-dumping, a rise in energy and raw materialprices and the increased competition from Asiatic emerging countries. Within this context,softwood lumber companies try to remain profitable and to maintain positive profit margins.

A softwood lumber firm can be considered as an MTS environment as its activities are drivenby forecasts. It is composed generally of multiple facilities including mills and distributioncenters. Unlike traditional manufacturing industries (i.e. assembly) which have a convergentproduct structure, the softwood lumber industry has complex transformation processes withheterogeneous raw materials (great diversity in terms of wood quality, diameters, length,

1. Natural Resources Canada, Forest products, accessed on September 15, 2015,http ://www.nrcan.gc.ca/forests/industry/13317

29

etc.), divergent product flows (generating many products at the same time) and radicallydifferent planning problems to be solved by each mill.

Since a high percentage of softwood lumber is used in the construction industry, demandfor lumber decreases in October—November and reaches a seasonal low during the wintermonths of December—February. Then, it experiences strong seasonal and cyclical rise in thesecond and third quarters. Prices are expected to move higher going into the summer asdemand increases. Thus, most seasonal fluctuations in softwood lumber prices can be ex-plained by demand seasonality related to construction activities. Although most of the timesawmills operate at full capacity, products are not always available in stock at the right timeto take advantage of price fluctuation for many reasons. First, there is almost no flexibility inraw material availability, depending on regulations of forestry activities and on the seasonalnature of harvesting operations, which limits the variation in the lumber sawing process. Se-cond, production operations are complex since divergent processes force different productsto be made dependently.

In this context, the dominant thinking currently in the Canadian lumber industry is to pro-duce the maximum volume with the available resource. Production is oriented towards largebatches to take advantage of economy of scale, resulting in large inventories, low flexibilityand low agility. A case study of a medium Canadian lumber firm, presented in the paper ofMarier et al. (2014), has shown that tactical planning such S&OP is important to take advan-tage of the cyclical nature of the softwood lumber industry. But in practice, S&OP is still notwell understood by such firms.

Furthermore, a large portfolio of softwood lumber products is offered to heterogeneous cus-tomers, having different attitudes and priorities. Dealers and distributors for example, aremore sensitive to price than to quality. Other customers, such as home improvement ware-house companies and housing component manufacturers, are willing to pay more for betterproducts and better services. RM is then interesting as a means to prioritize them, especiallysince a softwood lumber firm generally operates in supply-constrained environment as rawmaterial availability and capacity are bottlenecks. Consequently, all demand cannot alwaysbe fulfilled and the supply chain may offer fewer finished products than customer requests.So, sales managers are obliged to reject orders. Order promising based on RM concepts cansupport them to decide which orders should be rejected in anticipation of more valuableorders (Guhlich, Fleischmann, and Stolletz 2015), not only if not enough resources are avai-lable.

While decisions of accepting or refusing an order have to be near instantaneous and defini-tive, 3000 orders can be received weekly for a medium Canadian softwood lumber companywith three sawmills. Thus, sales managers are continuously confronted with the followingdecision problem : How can we synchronize mid-term, short-term and real-time sales de-

30

cisions –that have to be taken at multiple planning levels and at different frequencies– in away to maximize profits and enhance the service level offered to high-priority customers?

2.2.2 Related literature

Sales and operations planning (S&OP)

According to APICS Dictionary (2013), S&OP integrates all the business plans of a com-pany (supply, production, sales, customers, marketing, R&D and finance) in general terms,facilitates coordination between the various functions, and supports strategic and businessplans. Tuomikangas and Kaipia (2014) emphasize the tactical role of S&OP as a means oflinking company strategy and operational planning based on academic and practitioner li-terature. The S&OP process acts as a continuous mechanism that supports cross-functionalintegration (Oliva and Watson 2011). Despite the conflicting incentives in firms, S&OP faci-litates integrated supply chain planning and the involvement of all functions in every stagethrough a continuous criticism. Based on right information and effective planning proce-dures, a good performance can be achieved. S&OP can also support strategic decisions suchas capacity decisions (Olhager, Rudberg, and Wikner 2001). Moreover, S&OP supports inte-gration between the supply chains of different companies and ensures scheduling control toreduce delays (Affonso, Marcotte, and Grabot 2008). In an uncertain environment, S&OP ali-gns sales targets with resource availability. First, S&OP has an important role as a mediatorin improving operational performance in production environments characterized by marketincertitude (Olhager and Selldin (2007), Sodhi and Tang (2011), Feng et al. (2013)). By simu-lating an S&OP model with a stochastic demand, Feng, D' Amours, and Beauregard (2010)have proven that the S&OP process reduces effects of forecast errors in a Make-To-Orderenvironment. S&OP can also deal with order configuration uncertainty (Chen-Ritzo et al.2010). In contrast to the problems covered in these studies, our concern is not with S&OPperformance in different contexts, but with how S&OP can be integrated with RM concepts.

RM in production systems

Common studies in production systems have introduced RM concepts by different allocationmechanisms of the Available To Promise (ATP), which were summarized by Pibernik (2005).Standard ATP allocation mechanisms reject orders only if not enough resources are available,while in RM, due to the heterogeneity of customers, orders are also rejected in anticipationof more valuable orders (Guhlich, Fleischmann, and Stolletz 2015). Regarding application ofRM concepts in manufacturing context, two research streams can be distinguished. Withinthe first stream, the focus is on the implantation of allocation models in MTS context. Asecond stream has evolved from more advanced work on Assemble-To-Order (ATO) andMake-To-Order (MTO) environments, where both storable and non-storable resources areconsidered.

31

To the best of our knowledge, Meyr (2009) was the first to propose allocation models forMTS environments. He dealt with a deterministic demand and a known exogenous supplyand developed a linear programming formulation composed of two stages : "ATP allocation"and "real-time ATP consumption". This research was expanded by Azevedo, D' Amours, andRönnqvist (2016), who considers several mills and several products, while Meyr (2009) dealtwith just one mill and one product. The assumption of a deterministic demand may notbe applicable in some cases. So, Quante, Fleischmann, and Meyr (2009) considered demanduncertainty and proposed a dynamic programming formulation to take into account theimpact of consumption decisions. They showed that allocation model with nested bookinglimits always achieved better profits, like the deterministic model of Meyr (2009). UnlikeQuante, Fleischmann, and Meyr (2009) that assume that order reception period is the sameas the due date, our analysis considers a stochastic lead time. Pibernik and Yadav (2009)also dealt with stochastic demand, but the framework proposed takes into account carryover between allocation planning and order promising. This research was expanded (Samii,Pibernik, and Yadav 2011) to provide a formulation of a trade-off between the benefits ofreserving inventories for high-priority customers and the negative impact that this will haveon the overall system performance. These analyses were limited to a single period inventoryreservation problem of one product and just two classes of customers, in contrast to ourstudy which considers multiple demand classes, divergent product structure and multiplesourcing locations in a multi-period context.

Dynamic programming is often used by existing studies about RM in ATO and MTO produc-tion systems (Harris and Pinder (1995) and Gao, Xu, and Ball (2012)). Bid-price approachesare also commonly used as RM instruments for MTO production systems such as Spengler,Rehkopf, and Volling (2007) and Volling et al. (2011), but these analyses used fixed plan-ning horizons and a single plant case as opposed to the present paper in which a monthlyreplanning of several plants is considered over a year. Tsai and Wang (2009) is one of thefirst studies that considers more than one plant in an ATP mechanism for ATO productionsystem. Recently, Guhlich, Fleischmann, and Stolletz (2015) have developed a heuristic RMapproach using bid prices for a manufacturer using an ATO production system and facingstochastic demand. Stochastic approaches, such as in the Guhlich, Fleischmann, and Stolletz(2015) study, cannot be applied in our case since there is not enough data to model demandaccording to a known probability distribution.

All presented studies about RM application in production systems considered short-termplanning horizon. Within this time horizon, capacity levels cannot be extended and a knownexogenous supply is given. In particular, Azevedo, D' Amours, and Rönnqvist (2016) explai-ned how RM concepts can be introduced in a MTS context such as the softwood lumberindustry. However, the model that was proposed considered only a short-term planning ho-rizon and clearly ignored the potential profit that can be obtained by anticipating mid-term

32

demand. Our paper defines supply and production decisions as mid-term decision variablesin a context of divergent processes. So, we offer the possibility of taking advantage of de-mand and price fluctuations. In fact, an integrated demand management process is propo-sed in order to synchronize S&OP mid-term decisions with short-term and real-time salesdecisions taken according to RM concepts.

Order promising and medium term issues

Contrary to S&OP, order promising is a real-time problem. It is a critical task (Fleischmannand Meyr 2003), as it has impacts not only on company profitability and customer service le-vel in the short, medium and long term, but also has significant influence on scheduling andexecution of manufacturing and logistics activities (Pibernik and Yadav 2009). The relevanceof integrating order promising with tactical planning tasks was exhibited in a built-to-ordercontext by Volling and Spengler (2011), which explicitly model order promising and masterproduction scheduling as distinct, interdependent planning functions. Based on rolling ho-rizons, the analysis revealed the capacity of the integrated system to capture the impact ofproduction planning routines on the responsiveness and reliability of the order fulfillmentsystem and, vice versa, that of order promising decisions on the performance of productionplanning. More complex transformation processes with heterogeneous raw materials anddivergent product structure are considered in this paper. Besides, unlike our problem set-tings, the Volling and Spengler (2011) study is not concerned about market seasonality andorder/customer differentiation. The study of Dansereau et al. (2014) is a demonstration thatintegrating RM concepts at the tactical level can help to achieve better returns by providing abetter alignment of production with various market conditions. The model proposed explo-red the customer heterogeneity at the tactical level and optimized ATP quantities exclusivelyreserved to each customer segment. However, this is too rigid to be applied in practice. Ourpaper offers more flexibility by integrating a tactical model with RM using nested bookinglimits at real-time level and so higher-profitable segments can have access instantaneouslyto quantities reserved for lower-profitable segments. More flexibility is also guaranteed byoperating in a rolling horizon environment and by changing decisions of how confirmedorders have to be fulfilled after receiving each order and after each tactical planning.

The interaction between order promising decisions based on RM and customer relationshipmanagement, which focuses on medium-term horizon, was discussed in the case study ofOvchinnikov, Boulu-Reshef, and Pfeifer (2014). They developed a general dynamic modeland showed that trade-offs need to be made to benefit from low-value customers. A simi-lar point of view is discussed in our paper, based on S&OP decisions instead of customerrelationship management considerations.

33

Order reassignment

The idea of order reassignment can be related to the idea of flexible products presented inPetrick et al. (2012) and Gönsch, Koch, and Steinhardt (2014), in which the firm retains theright to specify later some of the details of a sold flexible product. In our case, decisionpostponement concerns the provenance from which an accepted order should be fulfilled.

2.3 Proposed demand management process

In this section, we describe the basic assumptions and the relevant decisions of the demandmanagement process that we propose to support sales decision making in firms such asCanadian softwood lumber firms.

Assumption 1 (Customers) : A company generally offers its products p to different marketsm, which refer to customers from different geographical regions (Azevedo, D' Amours, andRönnqvist 2016). Each market m can be split into customer segments g. Customer segmen-tation is a strategic task and a frequently applied tool in marketing science (Hofmann, Beck,and Füger 2013) to group the various types of customers and their behaviors and requi-rements, according to different criteria such as willingness to pay (Feng and Xiao (2000),Zhang, Wu, and Jin (2006), Li and Chen (2010)), quality sensitivity (Xiaodong et al. 2007),lead times (Li and Chen 2010), etc. This can provide the company with comprehensive infor-mation about its customers in order to identify sales opportunities (e.g. focus on profitable orloyal segments), to meet customer expectations and to follow segment evolution over time.

Assumption 2 (Demand information and time structure) : Sales and price forecasting are criticalinputs of the S&OP process (Mentzer, Myers, and Stank 2007). New information about de-mand and prices can be periodically obtained. While disaggregated forecasts can be madefor short-term horizon, medium-term forecasts are generally more dubious and aggregated.Forecasts aggregation (or disaggregation) can be applied to multiple dimensions simulta-neously : product families or single products, customer markets/segments or individualcustomers, different periods of time, etc. For instance, considering a medium-term horizoncomposed of T weeks, new forecasts of market demands Dmax

p,m,t and market prices αp,m,t canperiodically be available as shown in Figure 2.1.

Weekly market forecasts for short-term periods (weeks from the first week τ to week τ +T′ −

1, considering that the short-term horizon is composed of T′weeks as T

′<< T ) and monthly

market forecasts for medium-term periods (months from week τ + T′

to week τ + T − 1).Moreover, new weekly short-term forecasts of segment demands dmax

p,g,t and segment pricesβp,g,t can be available each month. S&OP process can be re-executed as soon as new forecastsare available.

34

FIGURE 2.1 – Available forecasts for short-term and medium-term horizons

Assumption 3 (Decisions) : Based on assumptions 1 and 2, firms looking to integrate S&OPand order promising based on RM concepts have four principal decision-making stages aspresented in Figure 2.2.

1. S&OP : Considering market demand forecasts Dmaxp,m,t and market prices forecasts αp,m,t,

contracts, sales commitments zsellp,g,t made in previous periods and current inventories

ip,n,0 in each node n ( a mill or a warehouse), we execute the S&OP every month overmedium-term horizon (e.g. twelve months) to predetermine supply, production, trans-port plans and market sales Vp,n,m,t for each product p expected to be sold from node nto market m at period t.

2. Allocation planning : Based on weekly short-term segment demand forecasts dmaxp,g,t and

segment prices forecasts βp,g,t, we allocate short-term market sales Vp,n,m,t to differentcustomer segments over short-term horizon (e.g. eight weeks). Commitments zsell

p,g,t tosell product p to segment g at period t, already made in previous periods, and weeklysegment forecasts dmax

p,g,t respectively represent lower and upper bounds for segment al-locations Xp,n,g,t(quantities of product p from node n allocated to segment g for periodt). In industrial practices, S&OP and allocation planning are mostly planned by dif-ferent teams. Nevertheless, it can be advantageous to simultaneously perform them assoon as we receive new forecasts (e.g. at the beginning of each month).

3. Booking limits identification : Before making promises, we identify, for each segmentg′

and for each period t′, from which allocations Xp,n,g,t we can consume by setting

booking limits (BLs) for each combination of segment g′and period t

′, based on expec-

ted profit margin ηn,g′ ,p,t,t′ for selling a product p, available in node n at period t, tosegment g

′at period t

′(t ≤ t

′). This stage will be detailed more in Section 2.4.2.

4. Real-time order promising and reassignment of orders to allocations : When we receive

35

a new order, we have to decide if we accept or refuse the order and from which alloca-tions Xp,n,g,t we should consume, considering BLs. We can also reassign previously ac-cepted orders, not yet delivered. Moreover, order reassignment has to be done monthlyafter each tactical planning.

FIGURE 2.2 – Proposed demand management process

36

Figure 2.2 illustrates the proposed demand management process. In this graphical repre-sentation, we suppose that S&OP is planned over a medium-term horizon (e.g. 12 months)and that we can make commitments just for short-term horizon (e.g. 8 weeks). Demand setby contracts needs to be satisfied and is considered as hard constraints in the S&OP. Thisdemand is included in incoming orders.

Assumption 4 (Orders) : Orders are treated individually. Batch order processing approachesin a similar setting as presented in the current paper are considered in Meyr (2009). We alsoassume that the decision of accepting or refusing an order has to be instantaneous and defi-nitive. However, order assignment to sourcing locations is temporary and may be changed.Partial fulfillment is not allowed, but an order can be fulfilled from different sourcing loca-tions. Sequence of high and low priority orders is not considered since they are randomlyreceived most of the time for the softwood lumber case, but can be examined in furtherresearch. Although the expected periodical demand is approximately known based on fore-casts, the exact ordering quantity varies randomly.

2.4 Model formulation

Figure 2.3 illustrates a supply network of a multi-site softwood company. In such a MTS en-vironment, a company has several nodes n (n ∈ N), representing sawmills and warehouses.Nodes can be supplied by different sources s (s ∈ S) and sell to various markets m (m ∈M)

composed of differentiated segments g (g ∈ G). Manufacturing plants are equipped withdifferent types of resources e (e ∈ E) enabling various activities a (a ∈ A). A node n issupplied by sources Sn (Sn ∈ S) and can execute activities An (An ∈ A). An activity a re-fers to a drying, planing or sawing recipe, so that the activity level can define amounts ofconsumed inputs and generated outputs. APp (APp ∈ A) are activities generating product p(p ∈ P), which can then be consumed by activities ACp (ACp ∈ A). Each product p can betransported on roads (n, n

′) ∈ Rop. The S&OP horizon is composed of T periods.

2.4.1 Tactical model (Stages 1 and 2)

At the beginning of each month, a tactical model simultaneously plans S&OP and allocationplanning. An S&OP network model for softwood company was proposed by Marier et al.(2014). It makes decisions related to supply, production, handling, transportation and salesin order to optimize the total company profit margin over a fixed horizon composed of Tperiods. Sales decisions are set by customer markets. We extended the model of Marier et al.(2014) in order to :

— work with a rolling horizon planning : We made several adaptations so that at eachnew tactical planning execution, decisions of the previous plan are integrated,

— allow order reassignment while respecting previous sales commitments,

37

FIGURE 2.3 – Supply network of a multi-site softwood company

— incorporate the allocation planning (stage 2 in Figure 2.2) : First, short-term sales deci-sions have to be allocated to different customer segments. Second, sales commitmentsand demand forecasts are set as lower and upper bounds for both market sales targetsand segment allocations.

Tables 2.1, 2.2 and 2.3 present respectively sets, parameters and decision variables involvedin the tactical model. Next, the objective function and constraints related to supply, transport,sales, inventory holding, production, flow balances, allocation and non-negativity will bedepicted.

38

TABLE 2.1 – Sets

Sets DescriptionA Activities aE Resource types e ∈ E ={sawing, drying or planing resources}M Markets mG Customer segments gGm Customer segments g of the market mN Nodes n(sawmills and warehouses)P Products pS Supply sources sSn Sources s supplying node n (Sn⊆S)An Activities executed in node n /An⊆AACp Activities consuming product p / ACp ⊆AAPp Activities generating product p / APp ⊆ARo Roads (n, n’) ∈ N×N∪MRop Roads (n, n’) ∈ Ro allowing transport of product p (Rop ⊆ Ro)

TABLE 2.2 – Parameters

Parameters DescriptionTimeτ First period of the planning horizon IndexT Length of medium-term horizon WeekT’ Length of short-term horizon WeekPrices/Costsαp,m,t Selling price of product p to market m during period t $/Qtyβp,g,t Selling price of product p to segment g during period t $/Qtycsup

s,n,t Supply cost from source s to node n during period t (purchase +transport)

$/Qty

cholp,n,t Holding cost of product p in node n during period t $/Qty

cproa,n,t Production cost of activity a ∈ An during period t $

ctrap,n,n′ ,t Transportation cost of product p on road (n, n

′) ∈ Rop during per-

iod t$/Qty

Supplyλmin

p,s,t , λmaxp,s,t Minimum [maximum] supply of product p from source s during

period t$/Qty

Λminp,s , Λmax

p,s Minimum [maximum] supply of product p from source s during theplanning horizon

Qty

ϕp,s Percentage of product p in a lot supplied from a source s %Transportσn,n′ Transportation delay from node n to node n’ Weekup,n,n′ ,t−σn,n′

Quantity of product p that started to be transported on road(n, n

′) ∈ Rop before the beginning of the current planning horizon

(t− σn,n′ < τ)

Qty

νminn,n′ ,t

, νmaxn,n′ ,t

Minimum [maximum] quantity transported on (n,n’)during periodt

Qty

SalesDmin

p,m,t Minimum demand to fulfill of product p for market m during periodt (contracts)

Qty

39

TABLE 2.2 – Parameters (continued)

Parameters DescriptionDmax

p,m,t Maximum demand to fulfill of product p for market m during periodt (market demand forecasts)

Qty

dmaxp,g,t Maximum demand to fulfill of product p for segment g during per-

iod t (segment demand forecasts)Qty

zsellp,g,t Commitments to sell product p to segment g during period t, al-

ready made in previous order promising cyclesQty

Inventoryip,n,0 Initial inventory of product p in node n Qtyimaxn,t Maximum inventory allowed in node n during period t Qty

Productionδe,a,n Capacity of resource type e used by activity a ∈ An Hr∆e,n,t Capacity of resource type e in node n during period t HrΦcon

a,p Quantity consumed by activity a to produce product p QtyΦpro

a,p Quantity of product p generated by activity a Qty

TABLE 2.3 – Decision variables

Decision variables DescriptionIp,n,t Inventory of product p in node n at the end of period t QtyLa,n,t Production level of activity a ∈An over period t QtyRs,n,t Quantity received from source s to node n during period t QtyRp,s,n,t Quantity of product p received from source s to node n during per-

iod tQty

Up,n,n′ ,t Quantity of product p transported on road (n,n’)∈Ropduring periodt

Qty

Vp,n,m,t Quantity of product p sold from node n to market m at period t QtyXp,n,g,t Quantity of product p from node n allocated to segment g for period

tQty

The objective function (Equation (2.1)) maximizes the total company profit margin over all Tperiods. The first two parts of Equation (2.1) compute the total selling revenue, consideringsegment allocations for short-term and market sales targets for medium-term. Then, costsof the whole planning horizon are subtracted. We first consider total supply cost, includingpurchase and transportation of raw materials costs. Second, production cost is set dependingon the activity level over the planning horizon. Next, inventory holding costs and transportcosts are depicted. Since reassignment is allowed (decisions of how confirmed orders haveto be fulfilled can be changed when the tactical model is re-executed), sales decisions tosegment g at period t already taken in previous order promising cycles will be included inthe allocations Xp,n,g,t of the new tactical model.

The objective function is subject to several sets of constraints described in text as follows :

40

(2.2) Products are supplied in predefined proportions depending onthe source.

(2.3) Supply limits from each source s per period.(2.4) Supply limits from each source for the whole planning horizon.(2.5) Transportation limits on each road over a period.(2.6) Quantity of each product p sold from a node n to a specific market

m at a period t is defined as quantities of this product transportedfrom the node n to the market m.

(2.7) Quantities sold to each market m must respect a minimum de-mand to fulfill. The maximum between market demand forecastsand sales commitments is set as an upper bound : when sales com-mitments exceed market demand forecasts, they should be consi-dered as maximum limits for short-term sales.

(2.8) A yearly inventory cycle.(2.9) Maximum inventory.(2.10) Maximum resources capacities.(2.11) Product flow balances : The inventory in a node n at the end of a

period t can be generalized as the inventory of the previous per-iod, plus the quantity received at the current period (consideredonly for raw materials), minus the quantity consumed by produc-tion activities over the current period, plus quantity generated byproduction activities over the current period, plus the differencebetween incoming and outgoing flows over the current period. In-coming quantities can include quantities that started to be trans-ported before the beginning of the current planning horizon τ.

(2.12) Product flow balances for the period τ.(2.13),(2.14) All variables are non-negative.

Allocation planning constraints need more explanation. Constraints (2.15) allocate sales tar-gets for market m to segments Gm. Allocations should exceed sales commitments zsell

p,g,t (leftpart of Equation (2.16)). Similarly to Equation (2.7), segment demand forecasts dmax

p,g,t shouldbe considered as maximum limits for short-term sales, except when quantities already com-mitted exceed forecasts. So, the maximum between segment demand forecasts dmax

p,g,t and salescommitments zsell

p,g,t is set as an upper bound for allocations (right part of Equation (2.16)).

Maximize ∑p∈P

∑n∈N

∑g∈G

τ+T′−1

∑t=τ

βp,g,tXp,n,g,t + ∑p∈P

∑n∈N

∑m∈M

τ+T−1

∑t=τ+T′

αp,m,tVp,n,m,t

− ∑n∈N

∑s∈Sn

τ+T−1

∑t=τ

csups,n,tRs,n,t − ∑

n∈N∑

a∈An

τ+T−1

∑t=τ

cproa,n,tLa,n,t

− ∑p∈P

∑n∈N

τ+T−1

∑t=τ

cholp,n,t Ip,n,t − ∑

p∈P∑

(n,n′ )∈Rop

τ+T−1

∑t=τ

ctrap,n,n′ ,t

Up,n,n′ ,t (2.1)

Supply constraints

Rp,s,n,t = ϕp,sRs,n,t ∀p ∈ P, ∀n ∈ N, ∀s ∈ Sn, t = τ..τ + T − 1 (2.2)

41

λminp,s,t ≤ ∑

n∈NRp,s,n,t ≤ λmax

p,s,t ∀p ∈ P, ∀s ∈ S, t = τ..τ + T − 1 (2.3)

Λminp,s ≤ ∑

n∈N

t=τ+T−1

∑t=τ

Rp,s,n,t ≤ Λmaxp,s ∀p ∈ P, ∀s ∈ S (2.4)

Transport constraints

νminn,n′ ,t

≤ ∑p∈P

Up,n,n′ ,t ≤ νmaxn,n′ ,t

∀(n, n′) ∈ Ro, t = τ..τ + T − 1 (2.5)

Sales constraints

Vp,n,m,t = ∑t−σn,m<τ

up,n,m,t−σn,m + ∑t−σn,m≥τ

Up,n,m,t−σn,m

∀p ∈ P, ∀n ∈ N, ∀s ∈M, t = τ..τ + T − 1 (2.6)

Dminp,m,t ≤ ∑

n∈NVp,n,m,t ≤ max

Dmaxp,m,t, ∑

g∈Gmzsell

p,g,t

∀p ∈ P, ∀n ∈ N, ∀s ∈M, t = τ..τ + T − 1 (2.7)

Inventory holding constraints

Ip,n,τ+T−1 = ip,n,0 ∀p ∈ P, ∀n ∈ N (2.8)

∑p∈P

Ip,n,t ≤ imaxn,t ∀n ∈ N, t = τ..τ + T − 1 (2.9)

Production constraints

∑a∈An

δe,a,nLn,a,t ≤ ∆e,n,t ∀(e, n) ∈ E×N, t = τ..τ + T − 1 (2.10)

Flow balances

Ip,n,t = Ip,n,t−1 + ∑n∈N

Rp,s,n,t − ∑a∈ACp

φconsa,p Ln,a,t + ∑

a∈APpφ

proda,p Ln,a,t − ∑

(n,n′ )∈Rop

Up,n,n′ ,t

+ ∑(n′ ,n)∈Rop

∑t−σ

n′ ,n<τ

up,n′ ,n,t−σn′ ,n

+ ∑(n′ ,n)∈Rop

∑t−σ

n′ ,n≥τ

Up,n′ ,n,t−σn′ ,n

∀p ∈ P, ∀n ∈ N, t = τ + 1..τ + T − 1 (2.11)

Ip,n,τ = ip,n,0 + ∑n∈N

Rp,s,n,τ − ∑a∈ACp

φconsa,p Ln,a,τ + ∑

a∈APpφ

proda,p Ln,a,τ − ∑

(n,n′ )∈Rop

Up,n,n′ ,τ

+ ∑(n′ ,n)∈Rop

∑σ

n′ ,n>0

up,n′ ,n,τ + ∑(n′ ,n)∈Rop

∑σ

n′ ,n=0

Up,n′ ,n,τ ∀p ∈ P, ∀n ∈ N (2.12)

42

Non-negativity constraints

La,n,t, Rs,n,t, Rp,s,n,t, Ip,n,t, Up,n,n′ ,t, Vp,n,m,t ≥ 0

∀a ∈ A, ∀p ∈ P, ∀s ∈ S, ∀n, n′ ∈ N, ∀m ∈M, t = τ..τ + T − 1 (2.13)

Xp,n,g,t ≥ 0 ∀p ∈ P, ∀n ∈ N, ∀g ∈ G, t = τ..τ + T′ − 1 (2.14)

Allocation constraints

Vp,n,m,t = ∑g∈Gm

Xp,n,g,t ∀p ∈ P, ∀n ∈ N, ∀m ∈M, t = τ..τ + T′ − 1 (2.15)

zsellp,g,t ≤ ∑

n∈NXp,n,g,t ≤ max

(dmax

p,g,t, zsellp,g,t

)∀p ∈ P, ∀g ∈ G, t = τ..τ + T

′ − 1 (2.16)

2.4.2 Order promising model based on RM concepts (Stages 3 and 4)

Once the tactical model is executed, we start to receive demand from customer segments fordifferent delivery periods. An order promising model is required to instantaneously makepromises to orders, while respecting the medium-term decisions and previous previous salescommitments. Since the model is based on RM concepts, we use nested booking limits to de-cide from which allocations we should consume to fulfill segment demands for each duedate. So, we have to assign demand required by segment g

′for delivery period t

′to alloca-

tions xp,n,g,t initially set to a segment g for delivery period t. Since it is an assignment pro-blem, we formulate it as a linear programming (LP) model. In contrast to Meyr (2009) andAzevedo, D' Amours, and Rönnqvist (2016) order promising models, our formulation allowsorder reassignment (i.e. changing decisions of how confirmed orders have to be fulfilled aslate as possible).

Table 2.4 describes additional sets, parameters and decision variables involved in the orderpromising model. An order required by segment g

′for delivery period t

′and fulfilled from

allocations xp,n,g,t has to be transported at t′ − σn,g′ . Thus, yp,n,g,g′ ,t,t′ represent quantities

already transported at current period j, while Yp,n,g,g′ ,t,t′ are not transported yet and can bemodified.

Customers in the softwood lumber context can be categorized according to their willingnessto pay (Azevedo, D' Amours, and Rönnqvist 2016). In addition, various studies of FORAC 2

research consortium (e.g. : Frayret et al. (2007), Lemieux et al. (2008)) affirm that softwoodlumber companies also have to handle sporadic customer orders, corresponding to a spotdemand from occasional customers offering low prices. These customers are referred to asthe spot segment g in the model.

2. FORAC research consortium works in collaboration with forest products industry stakeholders (compa-nies and government) in the province of Quebec and contributes since its launch in 2002 to the advancement ofresearch in the forest products industry, https ://www.forac.ulaval.ca/en/home/

43

TABLE 2.4 – Additional notation for the order promising model

Set Description

G Spot segment g (G = g ⊆ G)

Parametersj Current period Indexσn,g′ Transportation delay from node n to segment g

′Week

ηp,n,g′ ,t,t′ Profit margin for selling a product p, available in node n at per-iod t, to segment g

′at period t

′Qty

qp,g′ ,t′ Quantity of product p required by segment g′

for period t′

$/Qtyxp,n,g,t Quantity of product p from node n allocated by the tactical mo-

del to segment g for period tQty

yp,n,g,g′ ,t,t′ Quantity from allocation xp,n,g,t set for segment g′

at period t′

already transported (t′ − σn,g′ < j)

Qty

Decision variablesYp,n,g,g′ ,t,t′ Quantity from allocation xp,n,g,t consumed by segment g

′for

period t′

not transported yet (t′ − σn,g′ ≥ j)

Qty

Figure 2.4 represents an example where all transportation delays are set to zero. Assignmentsare illustrated as arcs between allocations and requested quantities. Since order reassign-ment is allowed, we can review these assignments as often as needed, i.e. after each tacticalplanning and whenever a new order is received.

Nested booking limits (NBL)

The concept of booking limits (BLs) is used to take advantage of customer heterogeneity andprofitability variation over time. According to Talluri and Van Ryzin (2004), setting BLs is away to control the availability of capacity. In our case, allocations xp,n,g,t represent capaci-ties in each node n designated to a segment g for a delivery period t. With NBL, capacities(allocations xp,n,g,t) designated to a combination (segment g, period t) can be sold to othercombinations generating better profits. It is as though capacities overlap in a hierarchicalmanner depending on the expected profit margin. Figure 2.4 shows that, to fulfill demandrequested by segment g

′and delivery period t

′, NBL allow us to consume from :

— allocations set to segment g′

for delivery period t′;

— unconsumed allocations set for previous delivery periods t(τ < t < j) ;

— allocations set to spot segment g for any delivery period (Quantities allocated to spotsegment can be consumed by any other segment) ;

— allocations set to segment g′

for future delivery period t preceding period t′

and gene-rating lower profit than being consumed at period t

′(j ≤ t < t

′, ηp,n,g′ ,t,t ≤ ηp,n,g′ ,t,t′ )

— allocations set to segment g different from g′

and g for any future delivery period tgenerating lower profit than being consumed by segment g

′at period t

′(j ≤ t <

t′, ηp,n,g,t,t ≤ ηp,n,g′ ,t,t′ )

44

FIGURE 2.4 – Allocations assignments to quantity requested by segment g′

for due date t′

(example where all transportation delays areset to zero)

45

Mathematical model

The goal of the order promising model is to maximize the objective function expressed byEquation (2.17), which is the short-term profit margin of fulfilling demand requested forperiods between the current period j and the end of the short term horizon (τ + T

′ − 1).

∑p∈P

∑n∈N

∑g∈G

∑g′∈G

τ+T′−1

∑t′=j+σ

n,g′

t′

∑t=τ

ηp,n,g′ ,t,t′ Yp,n,g,g′ ,t,t′ (2.17)

The model is subject to the following constraints : First, constraints (2.18) ensure that quanti-ties consumed from allocation xp,n,g,t set to a segment g for delivery period t will not exceedxp,n,g,t. This includes quantities yp,n,g,g′ ,t,t′ consumed by delivered orders that we can no lon-ger change (be reassigned), which is expressed for allocations that have been designated forpast periods of the allocation horizon by Equation (2.19), defined only if τ < j.

Allocation consumption

∑g′∈G

τ+T′−1

∑t′=t

t′≥j+σ

n,g′

Yp,n,g,g′ ,t,t′ ≤ xp,n,g,t

∀p ∈ P, ∀n ∈ N, ∀g ∈ G, t = j..τ + T′ − 1 (2.18)

∑g′∈G

τ+T′−1

∑t′=j+σ

n,g′

Yp,n,g,g′ ,t,t′ +

j+σn,g′−1

∑t′=t

yp,n,g,g′ ,t,t′

≤ xp,n,g,t

∀p ∈ P, ∀n ∈ N, ∀g ∈ G, t = τ..j− 1 de f ined i f τ < j (2.19)

Second, nested booking limits (NBL) constraints are expressed by constraints (2.20) and(2.21) : we force forbidden consumptions to be zero in order to avoid consumptions fromallocations set to more profitable segments and delivery periods (consumptions representedby forbidden arcs in Figure 2.4).

Forbidden consumptions for NBL

Yp,n,g,g,t,t′ = 0 ∀p ∈ P, ∀n ∈ N, ∀g ∈ G, t′= j + σn,g′ ..τ + T

′ − 1,

t = j..t′ − 1, ηp,n,g,t,t > ηp,n,g,t,t′ (2.20)

Yp,n,g,g′ ,t,t′ = 0 ∀p ∈ P, ∀n ∈ N, ∀g′ ∈ G, ∀g ∈ G\{g′ , g},

t′= j + σn,g′ ..τ + T

′ − 1, t = j..t′, ηp,n,g,t,t > ηp,n,g′ ,t,t′ (2.21)

46

Third, to guarantee previous commitments, additional constraints are expressed by Equation(2.22). Quantities consumed by segment g

′for delivery period t

′always have to be equal to

demand of segment g′

for period t′. Otherwise, the new order cannot be fulfilled. Finally,

constraints (2.23) assure that all variables are non-negative.

Respect of previous commitments

∑n∈N

∑g∈G

t′

∑t=τ

Yp,n,g,g′ ,t,t′ = qp,g′ ,t′

∀p ∈ P, ∀g′ ∈ G, t

′= j + σn,g′ ..τ + T

′ − 1 (2.22)

Non-negativity

Yp,n,g,g′ ,t,t′ ≥ 0

∀p ∈ P, ∀n ∈ N, ∀g, g′ ∈ G, t

′= j + σn,g′ ..τ + T

′ − 1, t = τ..t′

(2.23)

2.5 Data generation and experiments

2.5.1 Data generation and assumptions

In order to validate the proposed demand management process, an experimental case (seeTable 2.5) is considered based on softwood lumber manufacturers located in Eastern Canada.In this region, lumber manufacturers principally offer their products to Central Canadianmarket (CAC), Eastern Canadian market (CAE), Northeastern American market (US) and aspot market. There is little data available to model demand according to known probabilitydistribution. Generally, softwood firms only keep the information on shipped quantities,not demand information (Lemieux et al. 2008). Moreover, if a substitute product is shipped,the original demand information disappears from the database. Therefore, we made simpleassumptions to have market and segment demand and further ordering quantities, whichseem to fit the real case of most softwood lumber firms from the Eastern Canadian region.

TABLE 2.5 – Scope of the simulated case

Sets Size DetailsNodes (sawmills) 3Products 10 2x4 8’, 2x4 12’, 2x4 14’, 2x4 16’

2x6 8’, 2x6 12’, 2x6 14’, 2x6 16’Premium grade products

Markets 4 US, CAE, CAC and spot marketSegments 10 Spot market is composed of one segment. Other markets are

composed of 3 segments each.Average number of or-ders incoming weekly

100 Average weekly arrival rate is one order per combination (seg-ment, product), where one product is required per order.

47

For data generation, we assume that the yearly global demand is 150% of the maximumoutput that can be produced by pushing an infinity of supply into the supply chain. Thisassumption is acceptable for lumber softwood commodity products (Marier et al. 2014) sincedemand is too high and firms often produce at full capacity. Based on Frayret et al. (2007)and Lemieux et al. (2008), we assume that 80-95% of demand corresponds to a demand fromUS, CAE and CAC markets, while 5-20% of demand corresponds to a spot demand fromoccasional customers offering low prices (in our case, 0.8 of US market price). Such as in thestudy of Azevedo, D' Amours, and Rönnqvist (2016), we assume that markets US, CAE andCAC are composed of three segments each :

— High-priority customers (10% of the market demand), typically home improvementwarehouse companies and housing component manufacturers (Gaston and Robichaud2017), are ready to pay 10% more than the market price to have shorter transport lead-times,

— Medium-priority customers representing the majority of customers (70% of the marketdemand) pay exactly the market price,

— Low-priority customers (20% of the market demand), typically dealers and distributors(Gaston and Robichaud 2017), pay 10% less than the market price.

In what follows, we consider various demand scenarios as presented in Table 2.6. We assumethat the demand can be seasonal or stable and that the demand of the spot market (conside-red as one customer segment paying low prices) represents 20% or 5% of the total demand.Scenario 1 is the most realistic scenario for the softwood lumber context in Eastern Canada.Appendix A gives more details about how data have been generated.

TABLE 2.6 – Demand scenarios considered for data generation

Scenarios % of the demand of spot market Seasonal demand1 20% X2 20%3 5% X4 5%

2.5.2 Experiments

Experiments will be conducted in order to evaluate the benefits of integrating S&OP andRM concepts. Four demand management processes (see Table 2.7) will be evaluated withthe demand scenarios presented in Table 2.6. For the tactical level, we consider two differentlengths T of medium-term horizon : 8 and 52 weeks. With T = 8, the tactical model is usedjust for allocation planning, while with T = 52, we have S&OP coupled with allocationplanning. For the operational level, we compare two order promising models : the first modeluses NBL, while the second model handles orders according to a First-Come First-Served

48

TABLE 2.7 – Simulated demand management processes

Tactical level Operational/execution level Process Tactical constraints Operational constraints

S&OP/Allocation Nested Booking Limits (SA-NBL) 2.1-2.16 2.17-2.23

planning (T = 52) First-Come First-Served (SA-FCFS) 2.1-2.16 2.17-2.19, 2.22-2.23

Allocation Nested Booking Limits (A-NBL) 2.1-2.16 2.17-2.23

planning (T = 8) First-Come First-Served (A-FCFS) 2.1-2.16 2.17-2.19, 2.22-2.23

basis (FCFS) and decides if we accept or refuse each order assuming that all allocations areavailable to all (we do not use booking limits).

SA-NBL is the demand management process proposed in Figure 2.2, integrating S&OP, al-location planning and an order promising model using NBL. For all processes presented inTable 2.7, order reassignment is allowed (it can be done after receiving each order and aftereach tactical planning), and the tactical level is re-planned each month. The simulation isconducted with weekly planning periods over a year.

For each process presented in Table 2.7, a simulation algorithm developed in Visual Ba-sic.NET sequentially called the tactical and the order promising models. These models aredeveloped within IBM ILOG CPLEX Optimization Studio version 12.4. More details aboutthe simulation algorithm can be found in Appendix B. We needed 8.5 seconds for each or-der processing and so a total of almost 12 hours for all the orders of a year (8.5 sec/order x100 orders on average/week x 52 weeks). Expanding 8.5 seconds for each order processingseems to be acceptable in practice since in the worst case for a medium softwood lumberfirm, 600 orders will be received daily and so a total processing time less than 1.5 hours willbe needed.

Five replications 3 are simulated, i.e. a different seed is used for each replication to generatedifferent lists of orders as presented in Appendix A. We should note also that a warm-upperiod has been considered.

2.6 Results and discussion

2.6.1 Results analysis

In order to evaluate the global performance and the service level offered to high-prioritycustomers, results are analyzed regarding three performance indicators :

— The yearly profit margin (YPM) is calculated as the total selling price minus produc-

3. The number of replications was sufficient to observe a significant difference between the compared pro-cesses.

49

tion, transportation and inventory costs. This output is measured over a year to takeinto account the benefits of tactical planning considering cyclical rises of demand/price.

— The yearly sales (YS) represent the total volume sold and delivered over a year.

— The HP fill rate (HPFR) measures the proportion of demand received from high-prioritycustomers that has been fulfilled.

We evaluate the benefits of integrating S&OP and NBL compared to process A-FCFS, whichconsiders only a short-term allocation planning and makes real-time decisions accordingto FCFS basis. We use average values 4 of YPM, YS and HPFR to compute the benefits 5 ofintegrating S&OP and NBL in the different demand scenarios (see Table 2.8). Thus, we caninvestigate i) the benefits of the S&OP by comparing SA-FCFS to A-FCFS, ii) the benefits ofthe NBL by comparing A-NBL to A-FCFS and iii) the benefits of integrating both S&OP andNBL by comparing SA-NBL to A-FCFS.

The value of integrating S&OP and NBL is obvious in Table 2.8 : SA-NBL process achieves thehighest high-priority fill rates (HPFR) and an improvement of the yearly profit margin (YPM)ranging from 69% to 119%. The impact of integrating S&OP and NBL is more significantwith seasonal demand and with a low proportion of spot demand. In what follows, a moredetailed analysis is depicted.

TABLE 2.8 – Benefits of integrating S&OP and NBL compared to process A-FCFS

Spot demand Seasonality Processes YPM YS HPFR

20% X SA-NBL 84% 21% 65%SA-FCFS 49% 23% 19%

A-NBL 24% 0% 63%20% - SA-NBL 69% 21% 65%

SA-FCFS 46% 22% 17%A-NBL 22% 0% 62%

5% X SA-NBL 119% 25% 67%SA-FCFS 70% 26% 23%

A-NBL 18% 0% 64%5% - SA-NBL 103% 25% 68%

SA-FCFS 66% 25% 24%A-NBL 20% 0% 65%

Benefits on the yearly sales (YS) and the high-priority fill rate (HPFR)

Regarding the yearly sales (YS), we can see that we can sell 21-26% more by integratingS&OP with NBL/FCFS order promising models. Indeed, SA-NBL and SA-FCFS processes

4. Average values through the five replications.5. Example : YPM benefits for A-NBL=(Average value of YPM for A-NBL – Average value of YPM for A-

FCFS)× 100 ÷ (Average value of YPM for A-FCFS).

50

could deal better with the demand rise occurring at the mid term. Moreover, it is obviousthat the use of NBL does not affect the yearly sales (YS), but allows us to fulfill more high-priority demand and drives a more efficient use of the resources. As shown in Table 2.8, SA-NBL and A-NBL achieve an improvement around 65% of the high-priority fill rate (HPFR)compared to A-FCFS.

Benefits on the yearly profit margin (YPM)

YS and HPFR average values seem to be relatively stable through the different demand sce-narios that we have considered, in contrast to the yearly profit margin (YPM). Figure 2.5gives more details about the average values and the 95% confidence intervals of the YPM.

FIGURE 2.5 – Yearly profit margin

First, Figure 2.5 allows us to confirm findings of previous studies of Marier et al. (2014) andAzevedo, D' Amours, and Rönnqvist (2016) in the softwood lumber industry. In fact, we canconfirm statistically, based on 95% confidence intervals, that :

— we can achieve a better yearly profit margin by integrating S&OP : a benefit rangingfrom 46% to 70% can be observed if we compare SA-FCFS to A-FCFS,

— we can achieve a better yearly profit margin by using NBL : a benefit ranging from18% to 24% can be observed if we compare A-NBL to A-FCFS since A-NBL prioritizesorders from high-priority segments (an improvement of the HPFR up to 65% comparedto A-FCFS), while the A-FCFS does not differentiate between orders and focuses onfeasibility rather than profitability.

51

In addition, we can see that, as expected, the SA-NBL process is statistically 6 better thanother processes regarding the yearly profit margin (YPM) for all demand scenarios. Thevalue of integrating S&OP and NBL (SA-NBL compared to A-FCFS) can be reflected by abenefit on the YPM ranging from 69% to 119%. The higher profit generated by integratingS&OP and NBL is a result of the increased revenue by fulfilling more orders from profitablecustomers and for more remunerative periods.

Moreover, the SA-NBL process significantly takes advantage of demand seasonality. This canbe proved by comparing the YPM confidence intervals of the SA-NBL with stable demandvs. seasonal demand. This demonstrates that a potential profit can be obtained by efficientlyallocating the limited capacity (i.e. accumulating inventories to be sold when demand risesin high-price periods and rejecting orders, not only if not enough resources are available, butin anticipation of more valuable ones for more remunerative periods).

The impact of integrating S&OP and NBL is more significant if low-profitable demand suchas spot demand represents a low proportion of the total demand (Spot demand represents5% of the total demand in the left part of Figure 2.5). This underlines the interest of thecustomer segmentation : offering specific services (shorter lead-times in our case) to custo-mers less sensitive to price and preserving long-standing relations with them can potentiallysupport softwood lumber companies to remain profitable.

2.6.2 Managerial implications

In this study, firstly, we aim to appreciate the value of integrating revenue management (RM)and S&OP. Considering current demand management practices and existing IT-systems, wedeveloped a platform –integrating an S&OP model with an order promising model based onRM concepts– in order to help managers with the implementation challenges. Indeed, the in-tegrated demand management process proposed in Section 2.3 illustrates the different stagesof the integration and so supports managers having limited experience with RM\S&OP.

Second, we provide evidence to managers in regard of the benefits of implementing differentstrategies of integration. Our simulation results demonstrate, based on a case study in thesoftwood lumber industry, the improvement on the yearly profit margin and the service leveloffered to high-priority customers that can be achieved by integrated demand managementprocesses. This can help managers to overcome their fear of changes and losses, which is acommon barrier to the introduction of RM and S&OP evoked in literature (Kolisch and Zatta(2011), Oliva and Watson (2011), Noroozi and Wikner (2017)).

We should note that, in addition to the required models/softwares to implement the inte-grated demand management process proposed, we need the involvement of the different

6. The 95% confidence intervals of SA-NBL process do not overlap with the confidence intervals of otherprocesses

52

actors who participate in the process (Oliva and Watson (2011), Noroozi and Wikner (2017)).Hence, the realization of the benefits of integrating S&OP and RM is bound to the culturalcontext of the organization and requires cross-functional efforts (production, sales, distribu-tion, logistics, finance, marketing ...). Efforts should also be made to ensure the acceptanceof the RM on the client side since it implies the prioritization of high-priority orders.

The platform developed in this study can be also an efficient tool for softwood lumber ma-nagers to simulate new business models. It can be used to evaluate various what-if scena-rios and to anticipate how a demand management process integrating RM and S&OP willbe affected in typical situations like introducing a new high value-added product, chan-ging the capacity of sawmills, entering a new market, concluding contracts with other sup-pliers/customers, etc.

2.7 Conclusion and future research

In this paper, we extend the research in demand management for MTS manufacturing sys-tems. While existing studies dealt separately with revenue management (RM) and S&OP, wepropose a process integrating these two common methods and capturing feedbacks betweendifferent sales planning levels.

The proposed simulation framework offers guidance for a business problem presently facedby managers in softwood lumber industry and provides a deeper understanding of the linkbetween the S&OP and the order promising function, particularly when the organizationstrategy focuses on customer heterogeneity. Considering differentiated demand segments,divergent product structure and multiple sourcing locations in a multi-period context, wedevelop an order promising model using nested booking limits (NBL) and allowing orderreassignment, while respecting S&OP decisions and previous sales commitments. A rollinghorizon simulation is used to evaluate the performance of the integrated process in variousdemand scenarios.

Simulation results provide evidence of the value of integrating RM and S&OP and showthat we can offer better service level to high-priority customers and higher profit margin byintegrating S&OP and NBL compared to common demand management practices.

In this study, supply, production and transport decisions are limited to the aggregated tacti-cal level assuming that these optimal decisions can be implemented at the operational level.In reality, operational plans need to be taken into account. Also, considering different supplychain setups and different market variations may be of theoretical and practical interest.

53

Funding

The authors would like to thank FRQNT and NSERC for financial support.

54

Appendix A

Data generation

Weekly demand and prices are generated by a Visual Basic for Applications (VBA) code asfollows :

— The yearly global demand is 150% of the maximum output that can be produced bypushing an infinity of supply into the supply chain. Then, we compute average weeklydemand by dividing yearly demand by 52 weeks. We multiply the average weeklydemand by seasonality factors to obtain the weekly demand. Seasonality factors areequal to 1 with stable demand (for scenarios 2 and 4 presented in Table 2.8).

— Since US market represents the largest export market for Eastern Canadian softwoodcompanies, we set weekly demand forecasts of US market, CAE market, CAC marketand spot market as respectively 40%, 20%, 20% and 20% of weekly global demand forscenarios 1 and 2 and 40%, 27.5%, 27.5% and 5% of weekly global demand for scenarios3 and 4 (see Table 2.8).

— We suppose that, for each market, segments 1, 2 and 3 require respectively 10%, 70%and 20 % of all market demand. We consider these quantities as segment demand fo-recasts.

— We suppose also that CAE market, CAC market and spot market offer respectively 0.9,0.9 and 0.8 of US market price. These prices are used as market prices forecasts.

— For each market, segments 1, 2 and 3 offer respectively 1.1, 1 and 0.9 of the marketprice.

— Unit transport costs are proportional to distance between nodes and segments.

Afterward, we randomly generate orders using probability distribution as follows. Assu-ming that we receive 100 orders weekly, i.e. 1 order per combination (segment, product),we generate random variables for as many as we have orders per combination (segment,product) in a year. For each order of a combination (segment, product) :

55

— We generate reception period based on inter-arrival times, which follow a Poisson dis-tribution. Average weekly arrival rate of order depends on product required and oncustomer segment.

— We generate delivery delays following a triangular distribution. Maximum, averageand maximum delays are respectively set to 1, 3 and 4 periods for segments 2 and 3and to 1, 2 and 3 periods for segments 1. Then, we deduce delivery periods.

— We compute average quantity required by an order of a combination (segment, pro-duct) as weekly segment demand forecasts of the product demanded divided by theaverage weekly arrival rate. Quantity demanded by an order is then deduced as in-verse of normal distribution using as mean the average value previously obtained and0.1×mean as a standard deviation.

— We generate orders as a list ordered by reception date.

56

Appendix B

Simulation algorithm

The simulation algorithm presented by Figure B.1 is as follows :

1. Initialize the current period j to period 1. Since we are at the beginning of the month,we go to step 2.

2. Execute the tactical model. New allocations decisions xp,n,g,t are generated taking intoconsideration previous sales commitments yp,n,g,g′ ,t,t′ . Go to step 3.

3. Execute the order promising model. New sales commitments yp,n,g,g′ ,t,t′ are taken. Goto step 4.

4. If we have an order received at period j, update demand requested qp,g′ ,t′ , then go tostep 3. Otherwise, go to step 5.

5. Compute end-of-period inventory ip,n,j and increment current period (j← j + 1). If thenew period j is larger than to 52, stop. Otherwise, if the new period j is the first periodof the month (j mod 4=1), go to step 2. Otherwise, go to step 4.

57

FIGURE B.1 – Simulation algorithm

58

Chapitre 3

Configuration et évaluation d’unprocessus intégré de gestion de lademande via un plan de remplissaged’espace et la technique de krigeage

Cet article, intitulé "Configuration and evaluation of an integrated demand management processusing a space-filling design and Kriging metamodeling", a pour auteurs Maha Ben Ali, SophieD’Amours, Jonathan Gaudreault et Marc-André Carle. Il a été publié dans le journal interdis-ciplinaire "Operations Research Perspectives" en janvier 2018 et a gagné le prix du meilleurpapier dans la compétition "4th David Martell Student Paper Prize in Forestry" dans le cadrede la conférence de la société canadienne de recherche opérationnelle CORS2018. La versionprésentée dans cette thèse est identique à la version publiée.

59

Résumé

Cette étude vise à i) analyser le comportement d’un processus de gestion de la demandeintégrant la gestion des ventes et des opérations et la promesse de livraison dans un envi-ronnement de fabrication pour les stocks, et ii) à comparer différentes approches dans uncontexte de capacité limitée. Les études typiques sur les processus de gestion de la demandeproposent d’analyser un nombre restreint de paramètres du processus en les faisant varierun à la fois. Or, des analyses plus poussées peuvent être réalisées en utilisant les plans deremplissage d’espace et la technique de krigeage. Dans cet article, on propose de compa-rer deux configurations du processus intégré de gestion de la demande d’une entreprise dubois d’œuvre multisite. Dans la première configuration, les commandes sont traitées selonle concept du premier-arrivé premier-servi. La deuxième configuration, utilisant des limitesde réservation limitées, permet de donner plus de priorité aux commandes provenant desclients les moins sensibles aux prix et dûes pour les périodes les plus payantes. En consi-dérant des séquences variées d’arrivée des commandes, on génère des métamodèles de kri-geage permettant de capter les relations non linéaires entre quatre facteurs incontrôlables(soit, l’intensité de la demande, la précision des prévisions de la demande, l’hétérogénéitédes clients et la variabilité de la taille des commandes) et trois mesures de performance (leprofit annuel, les ventes annuelles et le niveau de satisfaction des clients prioritaires). Unplan de remplissage d’espace est utilisé afin de tenir compte de différentes situations dumarché. Notre analyse démontre le potentiel d’amélioration qu’on peut atteindre en sollici-tant les clients prioritaires à exprimer leurs besoins avant de traiter les commandes moinsprioritaires et en utilisant des limites de réservation imbriquées.

60

Abstract

Objective: This research aims to develop a basic understanding of a demand managementprocess integrating sales and operations planning (S&OP) and order promising in a Make-To-Stock environment and to compare different demand management policies with limitedcapacity. Contribution: Typical researches about demand management processes analyzefew system specifications or vary few potential factors one at a time. Yet, additional insightscan be obtained by employing a space-filling design and Kriging metamodeling for analy-sis. Methodology: We compare two configurations of the integrated demand managementprocess. While the First-Come First-Served concept is used at the order promising level forthe first configuration, the second configuration uses nested booking limits and gives ad-vantage to profitable customers and attractive periods. Considering various order arrivalsequences, we generate Kriging metamodels that best describe the nonlinear relationshipsbetween four environmental factors (demand intensity, demand forecast error, customer het-erogeneity and coefficient of variation) and three performance measures (yearly profit mar-gin, yearly sales and high-priority fill rate) for Canadian softwood lumber firms. Since oursimulation experiments are time-consuming, we employ a Latin hypercube design to ef-ficiently take into account different market situations. Results: Our analysis reveals thepotential to improve the performance of the demand management process if we know high-priority customers needs before fulfilling low-priority orders and if we use nested bookinglimits concept.

61

3.1 Introduction

3.1.1 Motivation and background

The Canadian softwood lumber industry is struggling to cope with certain challenges. Theindustry difficulties are mainly due to the increased cost of woody supply and reduced de-mand during the last decade (Government of Quebec 2016), coupled with the increased low-cost competition from emerging countries in Asia and Latin America (Teguia et al. 2017).Moreover, sawmills' profitability can be severely affected by fluctuations in the Canada-U.S.exchange rates and the numerous softwood lumber disputes between Canada and the Uni-ted States (Government of Quebec 2016).

Canadian softwood lumber companies have employed cost-cutting strategies to maintaincompetitiveness and profit margins (Teguia et al. 2017). However, they must be able to re-main profitable in situations where markets experience disturbances. This requires a deepe-ned understanding of the market side of the supply chain to take advantage of sales oppor-tunities (Gaston and Robichaud 2017), and an improvement of existing processes by usingreal-time monitoring systems as well as integrated planning systems (Favreau and Ristea2017).

This research is motivated by the need for Canadian softwood lumber firms operating ina supply-constrained environment and facing heterogeneous and seasonal market, to im-prove their demand management process and to anticipate how this process will perform indifferent situations. The dominant thinking currently in the Canadian lumber industry is toproduce maximal volume from the available resource, which is constrained by raw materialavailability and complexity of divergent production processes. Although sawmills operate atfull capacity most of the time, they do not take advantage of seasonal fluctuations of pricesand of the willingness of some customers to pay more for better products and better ser-vices. To this end, an integrated demand management process (IDMP) has been proposedby Ben Ali et al. (2014). They integrated sales and operations planning (S&OP) and orderpromising models, particularly those based on revenue management (RM) concepts.

The integration between RM and S&OP is not well understood either in theory or in prac-tice, particularly for Canadian softwood lumber firms. It is unclear how an IDMP, that canbe configured differently as presented in Ben Ali et al. (2014), can perform facing variousorder arrival sequences and market disturbances. In fact, Canadian softwood lumber mana-gers are confronted with different challenges such as a change of demand intensity, a rise ofdemand variability, poor accuracy of demand forecasts and increasingly heterogeneous cus-tomers. The simulation of the IDMP proposed by Ben Ali et al. (2014) offers the possibilityto experiment several demand management approaches and to measure the effect of theseenvironmental factors on the IDMP performance.

62

Searching for effects by varying factors one at a time is an ineffective means to estimatethe factor effects (Kleijnen et al. (2005), Montgomery (2009), Law and Kelton (2000)) since itimposes restrictions on the number of factors and the number of values that these factorscan take with a limited simulation budget, and so fails to consider nonlinear relationships.Using space-filling designs, and then Kriging metamodeling, is advantageous as an efficienttool with time-consuming simulation experiments to estimate factor effects on the IDMPperformance in different situations.

Our paper aims i) to develop a basic understanding of the IDMP proposed by Ben Ali et al.(2014) facing various order arrival sequences and taking various market disturbances intoaccount and ii) to compare different demand management policies. For these purposes,we have to identify : which factors are expected to have the most significant impacts onthe IDMP? and how can they affect the performance (improvement or deterioration and inwhich situations)?

3.1.2 Contributions and paper structure

Most multi-level decision processes and integrated decision-support systems in manufac-turing context are too complex to be evaluated analytically and so have to be studied bymeans of simulation before implementation. This paper addresses the need to evaluate theability of a multi-level decision process to face the different factors that could affect its per-formance. One of the main contributions of this paper is the novel procedure to experimentand to analyze the behavior of an integrated demand management process (IDMP) under avariety of scenarios : we employ a space-filling design and Kriging metamodeling to scan theeffects of some relevant market factors on the IDMP performance. To the best of our know-ledge, our study is among the few papers which use space-filling design and Kriging in arealistic supply chain setting, particularly to analyze factor effects and to compare differentdemand management approaches/practices. In addition, as motivated by an industrial pro-blem, the paper discusses the potential implications of this analysis for firms operating insupply-constrained environments, such as Canadian softwood firms.

The remainder of this paper is organized as follows. Section 3.2 presents the related litera-ture. In Section 3.3, we describe the industrial context. Section 3.4 exposes the performancemeasures, the factors considered in the experimentation and the experimental design. WhileSection 3.5 explains the different steps for data generation and analysis, Section 3.6 presentsthe analysis results and discusses managerial implications. Finally, concluding remarks andfurther research opportunities are provided in Section 3.7.

63

3.2 Related literature

3.2.1 S&OP and Revenue management in manufacturing

S&OP is a tactical process which supports cross-functional integration (Oliva and Watson2011) and links company strategy and operational planning (Olhager and Johansson (2012)and Tuomikangas and Kaipia (2014)). In fact, it is important to create a specific leadershipstyle and a culture in the organization to ensure integrated demand management and sup-ply chain planning. This required the involvement of all functions in each stage through acontinuous mechanism. However, the survey of Wagner, Ullrich, and Transchel (2014) showsthat organizations' current S&OP performance is underdeveloped and many improvementsare indispensable to concretize the alignment process. The lack of participants' commitmentand information reliability, the absence of cross-functional integration and a siloed cultureare the main barriers that jeopardize S&OP success (Pedroso, da Silva, and Tate 2016).

Although there are diverse researches available concerning S&OP implementation (Pedroso,da Silva, and Tate 2016), the role of S&OP as a powerful tool for reaching business targets ismostly absent from the current literature (Tuomikangas and Kaipia 2014). Moreover, syste-matic revues of Thomé et al. (2012) and Tuomikangas and Kaipia (2014) show that there isstill a need for more in-depth case studies with multiple perspectives to provide a deeper un-derstanding and guidelines for companies to manage the S&OP implementation challenges.In this context, this paper aims to provide a better understanding of the link between theS&OP and the order promising function, particularly when the organization strategy focuseson customer heterogeneity.

While S&OP makes mid-term decisions, order promising is a real-time problem which hasimpacts not only on company profitability and customer service level in the short, mediumand long term, but also has significant influence on scheduling and execution of manufactu-ring and logistics activities (Pibernik and Yadav 2009). When all demand cannot be fulfilled,introducing RM in order promising activity can be considered as a powerful tool ensuringhigher profitability and forging a stronger relationship with customers less sensitive to price(Stadtler and Kilger 2005) : order promising concerns how to manage capacity allocation,aggregately set by tactical planning, to different customers and introducing RM in orderpromising activity consists in protecting capacity reserved for each customer segment bydefining booking limits (Phillips 2005). Regarding application of RM concepts in manufac-turing context, two research streams can be distinguished. Within the first stream, the focusis on the implantation of RM in Make-To-Stock (MTS) context (Meyr (2009), Quante, Flei-schmann, and Meyr (2009), Azevedo, D' Amours, and Rönnqvist (2016)). A second streamhas evolved from more advanced work on Assemble-To-Order environment ( Tsai and Wang(2009), Gao, Xu, and Ball (2012) and Guhlich, Fleischmann, and Stolletz (2015)) and Make-To-Order environment (Spengler, Rehkopf, and Volling (2007) and Volling et al. (2011)).

64

The relevance of integrating order promising with tactical planning tasks was exhibited ina built-to-order context by Volling and Spengler (2011), who explicitly model order promi-sing and master production scheduling as distinct and interdependent planning functions.Ben Ali et al. (2014) have taken a further step forward by considering complex transforma-tion processes with heterogeneous raw materials and divergent product structure, mid-termmarket seasonality and customer differentiation.

Unlike existing studies which dealt separately with S&OP and RM in complex manufactu-ring situations (See Appendices C and D), Ben Ali et al. (2014) proposed an IDMP includingS&OP at the tactical level and real-time order promising based on RM concepts at the ope-rational/execution level (see Figure 3.1). This IDMP supports sales decisions in a way tomaximize profits and to enhance the service level offered to high-priority customers : First,considering demand and prices forecasts, sales commitments made in previous periods andcurrent inventories, S&OP is executed monthly over medium-term horizon to predeterminesupply, production, transport and sales plans, taking into account demand and prices sea-sonality. Second, real-time sales decisions have to be taken for each received order based onRM concepts, which offers the possibility of prioritizing orders from customers less sensitiveto price and for more profitable periods and to select the most profitable sourcing location.Our paper proposes going further by examining how such integrated process will performusing different demand management policies and facing different market situations.

Among all researches dealing with S&OP and RM in manufacturing context (See Appen-dices C and D), analysis by running only a single system specification or by varying somepotential factors one at a time were performed. Nonetheless, these tests can lead to differentconclusions if we make some changes in the factor settings. Using a conventional Design OfExperiments (DOE), additional insights can be gleaned with the same simulation budget.

FIGURE 3.1 – The integrated demand management process (IDMP) proposed by Ben Aliet al. (2014)

65

3.2.2 Conventional designs of experiments for simulation systems in supplychain settings

Factorial designs (full or fractional) are the most popular DOEs used in supply chain settings(See Appendix E), but the disadvantage of these designs is that the number of scenariosgrows exponentially when the number of factors or the number of factor levels increases.Taguchi (1987)'s designs are also widely common to identify robust decision factor settings.These designs are limited to main effects, which is usually too restrictive for simulation en-vironments (Kleijnen et al. 2005). Employing finer grids (more than two or three levels) forsome factors is important to view nonlinear relationships.

Space-filling designs, including Latin Hypercube Designs (LHD), make the samples moreuniformly spread in the experimental region (Cavazzuti 2013). They can be employed forcontinuous factors or discrete factors with a potentially large number of levels (Law 2015).These designs are more interesting for time-consuming experiments like ours. On one hand,they are efficient and flexible for analysis. On the other hand, they use an attractive samplingtechnique to provide data with few restrictions on factors and to cover large design spaces(Kleijnen et al. 2005).

3.2.3 Kriging metamodeling

Metamodeling is usually employed to analyze time-consuming simulation experiments. Theobjective is to represent the Input/Output (I/O) function implied by the underlying simula-tion model, and so predict outputs for new factor combinations, other than those simulated.In particular, Kriging (also called Gaussian process modeling) is typically used to developglobal metamodels (Law 2015) : "Kriging models are fitted to data that are obtained for lar-ger experimental areas than the areas used in low-order polynomial regression metamo-dels"(Kleijnen 2015). Kriging has traditionally been used for deterministic computer models.However, during recent years, the application of Kriging to outputs from stochastic simula-tion models, as is our case, has been explored by Kleijnen (2015). Simulation analysts oftenuse LHD to generate the I/O simulation data to which they fit a Kriging metamodel (Kleij-nen 2015).

3.3 Industrial context and case study

Market characteristics : Confronting various trade and economic pressures, Canadian soft-wood lumber companies try hard to remain profitable and to maintain positive profit mar-gins (Dufour 2007). In this context, our case study, illustrated by Figure 3.2, is inspired fromsoftwood lumber manufacturers located in Eastern Canada. In this region, lumber manufac-turers principally offer their products to different markets such as the Canadian market, theNortheastern American market, etc. A large portfolio of products is offered to heterogeneous

66

FIGURE 3.2 – The case study : A supply network of a multi-site softwood company

customers, having different attitudes and priorities. Home improvement warehouse compa-nies and housing component manufacturers, for example, are willing to pay more for shorterlead times and personalized services. Other customers, such as dealers and distributors, aremore sensitive to price.

Demand characteristics : In this study, we deal with ten lumber commodity products. Demandfor such products greatly exceeds supply offered by the company, as is usually the case forsoftwood lumber companies in Eastern Canada. In addition, prices are expected to movehigher going into some periods of the year. Most of these seasonal fluctuations in softwoodlumber prices can be explained by demand seasonality related to construction activities.

Sawmills/production characteristics : Sawmills can be considered as a MTS environment as itsactivities are driven by forecasts. Unlike traditional manufacturing (i.e. assembly) whichhas a convergent product structure, sawmills have complex transformation processes (i.e.sawing, drawing, planning) with heterogeneous raw materials (great diversity in terms ofwood quality, diameters, length, etc.), divergent product flows (generating many productsat the same time) and radically different planning problems to be solved by each mill.

Although sawmills operate most of the time at full capacity, products are not always avai-lable in stock at the right time to take advantage of price fluctuation for many reasons. First,there is little flexibility in raw material availability, depending on regulations of forestry ac-tivities and on the seasonal nature of harvesting operations, which limits the variation in

67

the lumber sawing process. Second, production operations are complex since divergent pro-cesses induce the production of multiple products simultaneously.

The studied network, illustrated by Figure 3.2, is composed from three sawmills with thesame capacity and dispersed over Quebec province. Sawmills can be supplied from twosources and sell to various markets (customers from different geographical regions and sowith different transport costs) composed of differentiated segments (customers classifiedaccording to their willingness to pay).

Actual situation : Whatever the market conditions, the dominant thinking of the Canadianlumber manufacturers is to produce the maximum volume from the available resource. Pro-duction is oriented towards large batches resulting in large inventories, low flexibility andlow agility. Ben Ali et al. (2014) have shown the potential profit that can be obtained by ta-king into account demand/price seasonality and by rejecting orders, not only if not enoughresources are available, but in anticipation of more valuable ones from profitable customersand for more attractive periods.

Based on multiple meetings with softwood lumber managers from the Eastern Canadianregion, we identified that they have two principal preoccupations : to maximize the profitmargin and the sales and to sell scarce products to the right customer (i.e. high-prioritycustomers) at the right time. Therefore, in what follows, we will consider profit margin, salesand high-priority fill rate as performance measures.

3.4 Experiments

In this study, we follow the procedure recommended by Montgomery (2009) for designingand analyzing experiments (see Figure 3.3). We have already recognized the problem andidentified the objectives of the experiments. Next, we have to define the performance mea-sures which reflect the system/process performance. Then, we have to set the list of factorsand the categories that they can take or the ranges over which these factors will be varied.Depending on the objectives of the experiments and the number and the nature of factors,we have to choose the type of experimental design.

68

FIGURE 3.3 – Procedure for designing and analyzing experiments (adapted fromMontgomery (2009))

3.4.1 Performance measures

Based on sales managers objectives in softwood lumber industry, we choose to analyze re-sults regarding three performance measures (see Figure 3.4) :

— The yearly profit margin (YPM) is calculated as the total selling price minus production,transportation and inventory costs. This output is measured over a year to take intoaccount the benefits of tactical planning considering cyclical rises of demand/price.

— The yearly sales (YS) represent the total volume sold and delivered over a year.

— The HP fill rate (HPFR) measures the proportion of demand received from high-priority(HP) customers that has been fulfilled.

While the two first indicators are oriented to evaluate global performance, the last one concernsthe service level offered to HP customers.

3.4.2 Factors

We have clustered the factors examined in this study into categorical decision factors andcontinuous environmental factors. In what follows, combinations of values for environmen-tal factors are called environmental scenarios. Table 3.1 and Figure 3.4 expose the factors andtheir associated categories or ranges.

69

FIGURE 3.4 – Performance measures and factors

TABLE 3.1 – Factors and their associated categories/ranges

Factor type Factors Notation Categories/Ranges # of Combina-tions

Categorical Order promising approach A NBL, FCFS * 6decision factors Order arrival sequence S ASC, RAND,DESC † combinationsContinuous Demand intensity I [1.25,1.75] 24environmental Demand forecast error E [-20%,+20%] environmentalfactors Customer heterogeneity H [+5%,+25%] scenarios ‡

Coefficient of variation V [0,1]

*. NBL : approach using Nested Booking Limits, FCFS : First-Come First-Served approach.

†. ASC : ascendant sequence, RAND : random sequence, DESC : descendant sequence.

‡. see Table S1.1

Decision factors

In this study, we assume that customer orders are treated individually and that the decisionof accepting or refusing an order has to be instantaneous and definitive. However, orderassignment to sourcing locations is temporary and may be changed. Partial fulfillment isnot allowed, but an order can be fulfilled from different sourcing locations. Although theexpected periodical demand is approximately known based on forecasts, the exact orderingquantity varies randomly.

In this context, two categorical decision factors affecting the system performances are iden-tified based on Ben Ali et al. (2014)'s study :

Order promising approach (A) : reflects how orders have to be fulfilled. Quantities to sell foreach customer segment at each period of the year are already set by the S&OP at the tacticallevel (see Figure 3.1).Then, for each received order, real-time sales decisions have to be taken.

70

For this purpose, different order promising approaches can be considered. First, we considera First-Come First-Served approach (FCFS), which simply decides if we accept or refuse eachorder, based only on resources availability. FCFS approach will be compared to a second ap-proach (NBL) based on RM concepts and using Nested Booking Limits. This approach can beapplied in a manufacturing setting in order to take advantage of customer heterogeneity andprofitability variation over time. According to Talluri and Van Ryzin (2004), setting bookinglimits is a way to control the availability of capacity. NBL approach can support managersin a supply-constrained environment, such as in the softwood lumber case, to decide whichorders should be rejected in anticipation of more valuable orders, not only if not enoughresources are available. Further on, with nesting, capacities overlap in a hierarchical mannerdepending on the expected profit margin, so that capacities initially designated to a specificcouple (customer segment, period) can be sold to other couples generating better profits.

Order arrival sequence (S) : reflects how orders arrive at order promising level. In this study, weconsider three arrival sequences : a random sequence (RAND) where orders from differentsegments are randomly received, an ascendant sequence (ASC) where orders are received inan ascending order of priority i.e. low-priority orders arrive first, and finally a descendant se-quence (DESC) where orders are received in a descending order of priority i.e. high-priorityorders arrive first. S can be considered as a decision factor since, in our industrial context,sales managers can stimulate HP customers to express their needs before dealing with low-priority orders.

Environmental factors

Environmental factors are uncontrollable in the real-world, but they are estimated and ap-proximately controlled for experimental purposes. Inspired from market disturbances confron-ted by Canadian softwood lumber managers and S&OP and RM literature (See AppendicesC and D), we select four relevant environmental factors. Each factor can take a numeric valuein a defined range.

Demand intensity (I) : is introduced at S&OP level and at order promising level. It representsthe percentage of the production capacity required to fulfill the demand (Forget et al. 2009).A demand intensity I equal to 1 has been estimated by pushing infinity of supply into thesupply chain and observing the maximum production output that can be produced (i.e. thecapacity). Then, we calculate demand as : Demand = I×Capacity. Since we are dealing withlimited capacity, we vary I between 1.25 and 1.75, similarly to Dumetz et al. (2015).

Demand forecast error (E) : is introduced at S&OP level. Similarly to Azevedo, D' Amours, andRönnqvist (2016), demand forecasts of all products in all weeks present an error E between-20% and +20 % in terms of demand volumes. Demand forecasts are upper bounds for salesplanned by S&OP, such as in Ben Ali et al. (2014), and are computed as : Demand f orecast =

71

(100 + E)%× Demand = (100 + E)%× I × Capacity.

Customer heterogeneity (H) : is introduced at order promising level and reflects the willingnessto pay of customer segments : high-priority segments are ready to pay H% more than themarket price, while low-priority segments pay H% less than the market price. Medium-priority segments represent the majority of customers and the price that they will pay isequal to market price.

Coefficient of variation (V) : reflects the demand variability such as in Quante, Fleischmann,and Meyr (2009) and is introduced at the order promising level. Order size is affected by astandard deviation = V × average order size, while the average order size is calculated asthe total demand (already affected by I) divided by a fixed number of received orders.

3.4.3 Experimental design

Figure 3.5 illustrates the experimental design. We consider the 6 combinations of the catego-rical decision factors. Each combination is simulated for m different environmental scenarios(i.e. combinations of the continuous environmental factors) generated using a Latin Hyper-cube Design (LHD).

For each environmental scenario i (where i denotes a LHD row, i = 1..m), we will have nmultiple outputs yir (where r denotes a replication, r = 1..n). Then, we will apply Kriging toyi, the average outputs for decision factor combination i across the n replications, similarlyto Law (2015). So, a total of 6×m× n runs will be performed. We consider m = 24 1 and theLHD is designed by JMP software (see Table S1.1 in the Supplementary Material S1).

FIGURE 3.5 – Experimental design

1. The total number of runs per replication is 6× m = 6× 24 = 144, which is equal to the number of runs(24 × 32 = 144) of a full factorial design with 4 two-level factors (A, I, H, V) and 2 three-level factors (S, E).However, much more information can be obtained through our design.

72

3.5 Data generation and analysis

The Supplementary Material S1 visually schematizes the different steps for data generationand analysis.

3.5.1 Generating data

Due to time considerations, we performed 3 replications 2. For each environmental scenario(i.e. combination of the environmental factors), we generate data as presented in Figure S1.1,which consists in :

1. generating data for the S&OP level,

2. generating data for the 3 replications of the order promising level :

2.1. generating a list of orders for each replication r(r = 1..3), using different pseudo-random numbers. In fact, randomness in our experiments concerns generatingorders for the order promising level and includes inter-arrival times, lead timesand quantities required by customer orders : (1) We assume that we receive, onaverage weekly, 200 orders per week, one at a time. In order to generate the inter-arrival times for a given couple (customer segment, product), we used a Poissonprocess with an arrival rate proportional to the demand of this specific (customersegment, product). (2) The delivery dates ( and so lead times) are set according tocustomer segments, i.e. on customer willingness to pay more for a shorter delay.Lead times follow a triangular distribution whose parameters are respectively setto (1, 2, 3) periods (weeks) for HP segments and (1, 3, 4) periods for other seg-ments. (3) The quantity required by an order, associated to a given couple (cus-tomer segment, product), follows a normal distribution. The mean of the distri-bution is calculated as the demand forecasts divided by the expected number ofreceived orders 3. Then, the mean is multiplied by the coefficient of variation V toinclude the standard deviation.

2.2. sorting this list differently to obtain 3 arrival sequences : an ascendant sequence(ASC), a random sequence (RAND) and a descendant sequence (DESC).

So, for each environmental scenario i and replication r, 3 final lists are obtained, sorted res-pectively by order of priority and by reception date. Common random numbers are used, sofor each replication r, the same seed is used to generate data for the m different environmen-tal scenarios.

2. This was sufficient to assess the variability of the performance measures since we obtained 95%confidence-interval half-lengths that are less than 10% of the average values. 10% is considered as a reasonablerelative error (Law 2015), especially tempered by the cost associated with the current number of replications.

3. For a couple (customer segment s, product p) : the average volume of an order is equal to the demandforecast for (s,p) divided by the expected number of orders for the period.

73

3.5.2 Performing the experiments

We use generated data to simulate the integrated demand management process (IDMP) withtwo different approaches A (NBL or FCFS) at the order promising level. The S&OP modeland the order promising model are formulated as linear programs (LP) and are developedwithin IBM ILOG CPLEX Optimization Studio version 12.4. In order to simulate the behaviorof the integrated demand management process, a rolling horizon simulation is conductedusing an algorithm developed in Visual Basic. NET, which sequentially called the S&OPmodel (executed each month) and the order promising model (executed for each receivedorder). We consider 200 orders/week. We need 8.5 seconds for each order processing, andso a total of 24 hours for all the orders of a year (8.5 sec/order x 200 orders /week x 52weeks). Since we have 6 decision factor combinations as explained in Section 3.4.3, a total of6×m× n = 6× 24× 3 = 432 runs are performed, 24 hours each.

3.5.3 Generating Kriging metamodels for average outputs

For each environmental scenario i (i = 1..m), we have n multiple outputs yir (where r denotesa replication, r = 1..n). So, we compute the average outputs yi = ∑n

r=1 yir/n for each decisionfactor combination. Then, we generate Kriging metamodels for average outputs yi, similarlyto Law (2015) p.677.

Kriging metamodels are constructed using "Gaussian process platform" of JMP software, topredict the evolution of the performance measures (i.e. outputs y) for new combinations ofthe environmental factors (demand intensity I, demand forecast error E, customer heteroge-neity H and coefficient of variation V). The Kriging makes two assumptions (Kleijnen 2015) :First, the model assumption is that yi, the average simulation output at input combinationi = (I, E, H, V), consists of a constant µ and an error term δi that is a stationary covarianceprocess with zero mean :

yi = µ + δi (3.1)

Second, the predictor assumption is that yi′ , the predictor at an arbitrary "new" input com-bination i

′, is a weighted linear combination of all the "old" output data yi at m already

simulated input combinations yi (i = 1..m) :

yi′ =m

∑i=1

λiyi (3.2)

To select the optimal weights in Equation (3.2), Kriging uses the "Best Linear Unbiased Pre-dictor" criterion, which minimizes the "Mean Squared Error" of the predictor y. For moredetails about Kriging metamodels, see Kleijnen (2015).

74

3.5.4 Statistical analysis of the data

Our results are analyzed regarding three performance measures (i.e. outputs), as presentedin Section 3.4.1 : the yearly profit margin (YPM), the yearly sales (YS) and the HP fill rate(HPFR). As mentioned in Section 3.3, we consider ten lumber commodity products 4. In whatfollows, we present performance measures for all products together since we are interestedin the overall process performance of the company. In the following section, we analyzethe impact of the decision factors and then the impact of the environmental factors usingresponse surfaces, prediction profilers and the analysis of variance (ANOVA).

3.6 Results and discussion

3.6.1 Impact of the decision factors (order promising approach A and orderarrival sequence S)

Figure 3.6 exhibits the performance measures of the two approaches A for different order ar-rival sequences S and the 95% confidence intervals on estimates over various environmentalscenarios : points in Figure 3.6 represent average outputs (see Section 3.5.3) for the different(I, E, H,V) combinations. The yearly profit margin (YPM) and the yearly sales (YS) are respec-tively expressed in millions of Canadian dollars (million$) and in million board-feet measure(MMFBM).

Less variation is seen for YS (the confidence intervals overlap in Figure 3.6b), clearly dueto the limited capacity of sawmills compared to the total demand. However, YPM and theHP fill rate (HPFR) are considerably sensitive to the approach A. Gaps between FCFS andNBL approaches are statistically significant regarding YPM and HPFR since the confidenceintervals do not overlap when we pass from blue side to red side in Figures 3.6a and 3.6crespectively. The gap is more pronounced if the HP orders arrived after low-priority orders(ASC sequence).

Regarding the order arrival sequence S, gaps between the three sequences in Figures 3.6aand 3.6c are statistically significant only for the FCFS approach. This means that, if we useFCFS, it is more interesting to receive HP orders early since we do not anticipate the arrivalof HP orders, in contrast to NBL. Further, Figures 3.6a and 3.6c exhibit that YPM and HPFRbehaviors for the three sequences are too close to be significantly different if we use NBL :we can say that with this approach, no order arrival sequence is preferable.

4. We have a divergent product structure, so it is not possible to produce the different products indepen-dently.

75

(a) Yearly profit margin (YPM) (b) Yearly sales (YS)

(c) HP fill rate (HPFR)

FIGURE 3.6 – Performance measures for different decision factor combinations

3.6.2 Impact of the environmental factors (demand intensity I, demand forecasterror E, customer heterogeneity H and coefficient of variation V)

Response surfaces, prediction profiles and ANOVA tables are drawn by "Gaussian processplatform" of JMP software, based on Kriging metamodels of the different performance mea-sures for each decision factor combination (A, S).

Response surfaces

We start our analysis by visualizing response surfaces to have an overview about the generaltrends of all performance measures throughout factor ranges. Figure 3.7 shows examples ofresponse surfaces for the effects of demand intensity (I) and demand forecast error (E) on theyearly profit margin (YPM) for FCFS and NBL approaches with H=10%, V=0.5 and randomarrival sequence.

It can be seen in Figure 3.7 that, for both approaches, YPM increases as I and/or E increases.This can be explained as follows : As mentioned in Section 3.4.2, demand forecasts are upperbounds for sales planned by S&OP and are computed as Demand f orecast = (100 + E)%×I × Capacity. If we increase demand forecasts by increasing the demand intensity I and/orthe forecast error E, the S&OP allocates more for remunerative periods (i.e. periods whenprices are high). Despite the fact that sawmills capacity cannot fulfill all demand (and so thetotal volume produced is almost the same), allocating more for remunerative periods enables

76

FIGURE 3.7 – Response surfaces for the effects of demand intensity (I) and demand forecasterror (E) on the yearly profit margin (YPM) for FCFS and NBL approaches (customer

heterogeneity H=10%, coefficient of variation V=0.5 and random arrival sequence)

our IDMP to accept more orders in these periods 5. In our context, additional inventory costsgenerated for example by a positive demand forecasts error (E > 0% compared to E = 0%)are compensated by additional revenues generated by selling more in remunerative periods.

As an example, Figures 3.8 and 3.9 present, respectively for NBL approach and FCFS ap-proach, the variation of sales and inventories over a year considering different demand fo-recast errors (E=0% and E=20%) with random arrival sequence, I=1.5, H=10% and V=0.5.When E passes from 0% to 20%, the number of accepted orders in remunerative periodspasses from 2537 to 2701 with NBL approach and from 2433 to 2572 with FCFS approach(equivalent to an increase of sales in remunerative periods by 8% for NBL approach and 6%for FCFS approach).

The IDMP using FCFS approach anticipates remunerative periods and sets limits for salesonly at the S&OP level. However, the IDMP using NBL additionally reserves quantities forHP orders since it sets limits for sales at the real-time level too. This explains the differencebetween NBL and FCFS curves respectively in Figures 3.8 and 3.9.

5. Example for a specific (product, customer segment) : Assuming that real weekly demand is 100 units and that2 orders are due for each week, order size varies around 50 units (the average size per order). Demand forecastsare 100 units if forecast error E = 0% and 120 units if E = 20%. S&OP weekly allocates 100 units if E = 0% and105 units if E = 20% since we have a limited capacity. Suppose that we receive the following order list : order 1of 45 units due for period t, order 2 of 50 units due for period t, order 3 of 50 units due for period t + 1 and order4 of 55 units due for period t + 1. The IDMP will accept only 3 orders if E = 0% and 4 orders if E = 20%.

77

FIGURE 3.8 – Variation of sales and inventories over a year considering different demandforecast errors (E=0% and E=20%), NBL approach, random arrival sequence, I=1.5, H=10%

and V=0.5

FIGURE 3.9 – Variation of sales and inventories over a year considering different demandforecast errors (E=0% and E=20%), FCFS approach, random arrival sequence, I=1.5, H=10%

and V=0.5

Prediction profilers

Prediction profilers offer the possibility to see how our prediction models change as wechange settings of individual factors and to find optimal settings for your factors regar-ding all our performance measures at the same time. These two-dimensional multivariateprofilers are interesting in order to interact with responses, which is another benefit of usingmulti-factor Kriging metamodels. Figure 3.10 shows examples of the JMP profiler tool, whichpresent the response for each performance measure as it relates to each factor, i.e. how thepredicted response changes as one factor is changed while the others are held constant atthe current values of factors. Current values of factors (I=1.5, E=0%, H=10% and V=0.5) andcurrent predicted values of responses are presented in red respectively in the x-axis and they-axis.

78

FIGURE 3.10 – JMP profiler tool for I=1.5, E=0, H=10% and V=0.5

The first thing to note is that the HP fill rate (HPFR) varies considerably when we changethe order promising approach A (when we pass from left to right side in Figure 3.10). Regar-ding the environmental factors, HPFR is almost only sensitive to the coefficient of variationV (see 3rd row of each profiler). Indeed, HPFR declines when V goes over 0.6 (the bottomright corner of each profiler). Even so, we can generalize that, no matter the environmentalconditions, NBL has to be chosen if our objective is to improve HP customers' satisfaction.

Regarding the yearly profit margin (YPM) and the yearly sales (YS), the prediction profilersin Figure 3.10 confirm the important effect of the decision factors (order promising approachA and order arrival sequence S), especially for YPM. They also lead us to believe that themost pertinent environmental factors affecting the YPM are the demand intensity I and thedemand forecast error E (the four upper left squares of each profiler), and that YPM and YS

79

increase if I and/or E increases. The yearly sales (YS) are however stabilizing when E goesover +5% due to the limited production capacity. We can see also that, for the ascendantsequence ASC and with NBL approach, the yearly profit margin (YPM) declines when I goesover 1.7, due to the rise in quantities stocked and reserved for HP customers (arriving last insequence ASC). Finally, we note that the customer heterogeneity H significantly affects theYPM and YS if we use FCFS approach (the two upper squares in the 3rd column of each FCFSprofiler).

Analysis of variance (ANOVA)

Kriging metamodels may also be analyzed through ANOVA (Kleijnen 2015), which allows usto quantify/measure the effects already shown by response surfaces and prediction profilersas trends, so we can identify the most pertinent factors in different situations : the objectivefor this analysis is to examine the contributions of environmental factors and interactionsfor each decision factor combination (A,S). We can assume normal distributions, and so useANOVA, only for the yearly profit margin (YPM) and the HP fill rate (HPFR) 6. Conside-ring a significant contribution if it exceeds 10%, ANOVA tables presented in SupplementaryMaterial S3 give evidence that :

— The coefficient of variation V is the most pertinent environmental factor affecting theHPFR. In fact, the more V increases, the more large-size HP orders we have ; so it ismore often that a HP order will be rejected since backorders and partial fulfillment (i.e.to fulfill just a part of the order) are not allowed in our simulation settings.

— For both NBL and FCFS approaches, the demand intensity I and the demand forecasterror E represent a significant part of the contribution for YPM (a total of 86-96% forNBL and 65-85% for FCFS).

— The customer heterogeneity H represents a significant part of the contribution onlyfor YPM if we use FCFS approach. However, the earlier HP orders arrive, the less theYPM will be penalized by H (H contribution is 32%, 19% and 7% respectively for ASC,RAND and DESC sequences). In fact, since FCFS approach focuses on feasibility ratherthan profitability, it does not anticipate receiving more valuable orders. So, capacitycan be exhausted by less profitable orders (paying H% less than the market price) andcannot fulfill more profitable orders received later.

— There is no significant interaction between environmental factors.

3.6.3 Managerial implications

Our study suggests implications for both supply chain management researchers and practi-tioners. For supply chain management researchers, our paper provides an evaluation of the

6. See normality tests in Supplementary Material S2.

80

value of integrating two common concepts in the demand management research, namelyS&OP and revenue management (RM). In addition, we employ a space-filling design andKriging metamodeling, which is a relatively new procedure for realistic supply chain ma-nagement experiments. For practitioners, this paper provides a tool to evaluate the perfor-mance of an integrated demand management process (IDMP) in different industrial settings.The methodology proposed can support a sales manager to decide which configuration willbe appropriate depending on his specific context and to identify the actions to be conductedin order to improve the performance of the demand management process.

Our analysis demonstrates that in a supply-constrained environment such as the Canadiansoftwood lumber industry, managers can achieve better performances by integrating S&OPand RM : the IDMP makes an implicit trade-off between the objectives of the productionteam (dealing with divergent production challenges) and the incentives of sales team (foste-ring better relationships with profitable customers).

Facing the potential market disturbances, production/sales managers need to be supportedby a tactical plan such as S&OP in order to capture the possible revenue increase, rather thanproducing in a push mode. Our study demonstrates that sawmills should take advantage ofany rise of demand intensity by allocating more for remunerative periods, which is possiblewithin tactical planning and medium-term forecasting. In fact, forecasts are critical inputsto S&OP : demand and price forecasting plays a determining role in the overall planningactivities of a firm, especially in the forest industry (Feng, D' Amours, and Beauregard 2010)since forest product prices and demand are well known for their fluctuations. Moreover, ouranalysis has asserted that the performance of the IDMP is less affected by the forecastingerror if we use an order promising approach considering nested booking limits (NBL).

This study shows that NBL order promising approach is efficient to capture orders fromprofitable customers and for more remunerative periods, and so immunize the demand ma-nagement process against different environmental disturbances. However, the use of NBLrequires a deep understanding of the market. In fact, customer segmentation is needed togroup the various types of customers and their behaviors and requirements, according todifferent criteria such as the willingness to pay, loyalty, etc. In this study, we assume thatsome customers are ready to pay more to have shorter transport lead-times. Potentially, othervalue-added services can be considered, like the stability of product quality and partnershipagreements (see Lehoux et al. (2010)'s study in the pulp and paper industry). Customer Re-lationship Management initiatives can be used to identify customer segments and to reachthe customers who are most receptive to the products and services offered.

Considering current practices and existing IT-systems, managers can face challenges to im-plement RM and S&OP. Our results are illuminating interesting managerial practices thatcan be easily introduced before RM and S&OP implementation. We demonstrate that the

81

order arrival sequence should be taken into consideration : even if orders are fulfilled on aFCFS basis, sales managers in softwood lumber industry should start by stimulating high-priority (HP) customers to express their needs before dealing with low-priority customers toimprove the performance of the company.

3.7 Conclusion and further research opportunities

This paper aims to contribute to the research in demand management for MTS manufactu-ring systems and to analyze a process integrating S&OP and order promising, consideringdifferentiated demand segments, divergent product structure and facing various market dis-turbances. For these purposes, we use relatively novel techniques – a space-filling design andKriging metamodeling – in supply chain settings. We are also among the first who addressthe impact of decision and environmental factors on performances of an integrated demandmanagement process.

Our simulation results affirm that NBL approach can be a powerful tool to maximize reve-nues facing different environmental conditions. We also show how order arrival sequencecan play a relevant role, especially with high customer heterogeneity. Therefore, sales ma-nagers in softwood lumber industry should, first of all, intensify their efforts to know, asearly as possible, the needs of HP customers and to improve the performance of CustomerRelationship Management, which might be simpler than implanting a new demand mana-gement platform. Then, they should focus on customer heterogeneity by using an integrateddemand management process able to anticipate orders from profitable customers and formore remunerative periods.

It is important to note that the validation of experiments was done only for a specific in-dustrial case study. For generalizing, we provide the tool and the methodology needed toperform other simulation experiments with different settings and in other industry sectors,especially those dealing with stochastic behaviors in terms of supply, demand and manufac-turing operations and divergent production processes.

Future research efforts concerning the integrated demand management process validationmay provide some new insights. First, in other contexts, it could be interesting to includeother decision and environmental factors. Second, since in practice, prices offered for up-coming periods are uncertain at the order promising level, a scenario-based stochastic pro-gramming model could be considered at the tactical level. Third, other order promising op-tions such as partial fulfillment and substitution could be investigated.

82

Funding

This work was supported by the Natural Sciences and Engineering Research Council of Ca-nada (NSERC).

83

Appendix C

Analyzed factors in sales and operations planning(S&OP) literature

84

TABLE C.1 – S&OP literature

Factors Description Chen-Ritzoet al. (2010)

Feng,D' Amours,

andBeauregard

(2008)

Feng,D' Amours,

andBeauregard

(2010)

Hahn andKuhn(2012)

Lim,Alpan, andPenz (2014)

Sodhi andTang (2011)

Wochneret al. (2016) Total

Capacity flexibility or demand in-tensity compared to capacity

e.g. overtime hours, stock margins, productionpolicies × × × × × 5

Integrated / decoupled approaches × × × 3

Forecast errors Overestimation or underestimation of demandvolumes × × 2

Supplier flexibility e.g. Delays flexibility Emergency supplies × × 2

Operational costs e.g. production cost, unit purchase cost, unitshipping cost, unit raw materials cost × × 2

Market price × 1

Demand pattern e.g. gradually demand increase or demandpeak in a specific period × 1

Maturity of rework Rework rates and rework times × 1

Order flexibility rate Possibility to delay orders × 1

85

Appendix D

Analyzed factors in literature about revenuemanagement(RM) in manufacturing

86

TABLE D.1 – RM in manufacturing literature

Factors Description

Azevedo,D' Amours,

andRönnqvist

(2016)

Chevalieret al. (2015)

Chiang andWu (2011)

Gönsch,Koch, andSteinhardt

(2014)

Guhlich,Fleisch-

mann, andStolletz(2015)

Kim andBell (2015)

Ovchinnikov,Boulu-

Reshef, andPfeifer(2014)

Petricket al. (2012)

Pibernikand Yadav

(2009)

Quante,Fleisch-

mann, andMeyr(2009)

Raza (2015) Total

Productionshortage or de-mand intensity

Percentage of production ca-pacity required to fulfill thedemand

× × × × × × × 7

Demand varia-bility

Coefficient of variation or dif-ferent demand distributions × × × × × 5

Profit structureor heterogeneity

Difference of selling prices of-fered by different customersegments

× × × × 4

Forecast errors Overestimation or underesti-mation of demand volumes × × × 3

Order sizestructure

Vary the number of or-ders/Consider different ordersize for each segment

× × 2

Demand struc-ture

Difference between demandrates of different customersegments

× × 2

Lead time struc-ture

Difference between lead timeoffered to different customersegments

× 1

Network struc-ture

Compare parallel networkstructures and hub-and-spoke networks

× 1

Optimizationfrequency × 1

Number of pro-duct per order × 1

Flexible pro-ducts

Consider or not flexible pro-ducts × 1

Customer life-time value

Consider or not customer life-time value calculation × 1

Demand arrivalpatterns

Demand with no peak, de-mand with an early peak, de-mand with a middle peak

× 1

87

Appendix E

Recent literature of conventional DOE for simulationsystems in supply chain settings

TABLE E.1 – Conventional DOE for simulation systems in supply chain settings

Paper Research topic DOE type Objective*Bottani and Montanari (2010) Inventory management Full factorial 1Sandhu, Helo, and Kristianto (2013) Inventory management and information sharing Full factorial 1Nedaei and Mahlooji (2014) Supply chain scheduling Full factorial 1Bandaly, Satir, and Shanker (2014) Supply chain risk management Full factorial 1Dev, Shankar, and Debnath (2014) Inventory management and risk management Taguchi 1Ciancimino et al. (2012) Supply chain collaboration Latin Square 1Dominguez, Cannella, and Framinan(2015) Supply chain structure Full factorial 1,2

Hussain, Khan, and Sabir (2016) Inventory management Taguchi 1,2Ponte et al. (2016) Supply chain collaboration Fractional factorial 1,3Santa-Eulalia et al. (2011) Tactical planning and production control policies Taguchi 2Azadeh, Zarrin, and Salehi (2016) Supplier selection in a closed loop supply chain Taguchi 2Shi et al. (2013) Cross-docking distribution Full factorial + LHD 2Assarzadegan and Rasti-Barzoki (2016) Supply chain scheduling problem Full factorial 3Olaitan and Geraghty (2013) Production control LHD 2,3

* 1 :Developing a basic understanding of a simulation model/system, 2 :Finding robust decisions, 3 :Comparing the merits of various decisions/policies.

88

Appendix F

Supplementary Materials

S1 Steps for data generation and analysis

First, we start by generating 24 environmental scenarios (see Table S1.1) using Latin hyper-cube design with JMPs experimental design software.

Second, for each of the 24 environmental scenarios :

1. we generate data as presented in Figure S1.1, which consists in :

1.1. generating data for the S&OP level

1.2. generating data for 3 replications of the order promising level as follows :

1.2.1. generating a list of orders for each replication r, r= 1..3

1.2.2. sorting this list differently to have 3 arrival sequences :an ascendant sequenceASC, a random sequence RAND and a descendant sequence DESC.

2. we use data generated in step 1 to simulate the demand management process with 2different demand management approaches (NBL or FCFS). For the example presentedin the second column of Figure S1.1, S&OP data and the list of orders corresponding toreplication 1 and ASC sequence are used.

3. we compute the average values of the performance measures over the 3 replications.In the third column of Figure S1.1, we present the example for the ASC sequence andboth the NBL and the FCFS approaches.

89

TABLE S1.1 – Environmental scenarios

Environmental scenario I E(%) H(%) V1 1.47 18.26 5.87 0.092 1.75 -2.61 11.52 0.483 1.27 16.52 8.91 0.354 1.34 -11.30 12.83 0.175 1.45 14.78 7.61 0.916 1.36 -16.52 8.48 0.527 1.64 -20.00 11.96 0.228 1.55 -4.35 5.43 0.879 1.25 -13.04 13.70 0.7010 1.42 -7.83 10.22 1.0011 1.66 -18.26 9.78 0.7812 1.53 0.87 11.09 0.0013 1.60 20.00 10.65 0.3014 1.58 13.04 15.00 0.6115 1.29 4.35 5.00 0.6516 1.62 -14.78 6.30 0.3917 1.32 7.83 12.39 0.7418 1.49 2.61 8.04 0.4319 1.73 6.09 7.17 0.1320 1.51 -6.09 13.26 0.5721 1.40 -9.57 6.74 0.0422 1.71 11.30 9.35 0.8323 1.38 9.57 14.57 0.2624 1.68 -0.87 14.13 0.96

90

FIGURE S1.1 – Steps for data generation and analysis (1)

91

Finally, for each of the 6 decision factor combinations (see Figure S1.2) :

4. we generate a kriging metamodel based on the average outputs computed for eachenvironmental scenario as described step 3.

5. we analyze results and conclude by :

5.1. analyzing the impact of decision factors

5.2. analyzing the impact of environmental factors using response surfaces, predictionprofilers and ANOVA.

FIGURE S1.2 – Steps for data generation and analysis (2)

92

S2 Normality tests

In four of the six cases (Figures S2.1-S2.6), we cannot consider normal distribution for YSresponses. However, we can assume normality of YPM and HPFR responses in all cases.

FIGURE S2.1 – Normality tests for FCFS approach and ASC sequence

FIGURE S2.2 – Normality tests for FCFS approach and RAND sequence

93

FIGURE S2.3 – Normality tests for FCFS approach and DESC sequence

FIGURE S2.4 – Normality tests for NBL approach and ASC sequence

94

FIGURE S2.5 – Normality tests for NBL approach and RAND sequence

FIGURE S2.6 – Normality tests for NBL approach and DESC sequence

S3 ANOVA tables for YPM and HPFR

95

FIGURE S3.1 – ANOVA for FCFS approach

96

FIGURE S3.2 – ANOVA for NBL approach

97

Chapitre 4

Simulation d’une approche intégrée degestion des revenus pour un systèmede coproduction avec substitution deproduits

Cet article, intitulé "Simulating an integrated revenue management approach for a coproductionsystem with product substitution", a pour auteurs Maha Ben Ali, Sophie D’Amours, JonathanGaudreault et Marc-André Carle. La version présentée dans cette thèse est une version éten-due de la version acceptée à la conférence "Winter Simulation Conference 2018" en juin 2018.

98

Résumé

Cette étude est motivée par le cas du bois d’œuvre canadien où la substitution des produitsreprésente une pratique commune de satisfaction de la demande. On considère la substitu-tion comme une politique que l’entreprise pourrait adopter. Les articles existants sur l’appli-cation des concepts de gestion des revenus dans un contexte de substitution de produits sesont plutôt intéressés aux enjeux de fixation des prix et ont assumé que la substitution est unedécision du client en réaction aux écarts des prix entre les différents produits substituables.Dans cette étude, on propose un modèle générique de satisfaction de la demande intégrantla substitution des produits, ainsi que le contrôle de la capacité via les limites de réserva-tion imbriquées. Par la suite, on évalue la performance du modèle intégré face à différentsscénarios en le comparant à des modèles conventionnels de satisfaction de la demande. Par-ticulièrement, nous nous sommes intéressés à l’effet de proposer à certains clients privilégiésun produit de meilleure qualité au prix du produit original demandé, soit l’équivalent d’unsur-classement ("upgrading") pour les entreprises de service. Cette étude contribue à la litté-rature existante sur la planification hiérarchique. En effet, différents modèles de satisfactionde la demande sont intégrés avec un modèle de planification des ventes et des opérationsau niveau tactique, ce qui offre une visibilité à moyen terme pour les plans de production etd’approvisionnement. Une simulation en horizon roulant est menée afin de mettre en évi-dence les bénéfices d’intégrer la gestion des revenus et le sur-classement dans un contextede capacité limitée et de coproduction.

99

Abstract

Most revenue management publications dealing with substitutable products in a manufac-turing context have focused on pricing issues. They consider that substitution is a customer’sdecision which occurs as a response to product price differences. In our study, substitutionis considered as the firm’s policy. We focus on the extension of the revenue managementto practical applications in manufacturing and we are motivated by the Canadian softwoodlumber case where product substitution is a common demand fulfillment practice. We aim,first, to propose a generic consumption model integrating both capacity control and productsubstitution decisions and second, to evaluate the performance of this integrated model indifferent settings compared to other common demand fulfillment approaches. In additionto practical implications, our study contributes to the existing demand fulfillment literaturesince we integrate different consumption models, executed at the real-time level, with a Salesand Operations planning (S&OP) model. This offers additionally a medium-term visibilityto make supply decisions. A rolling horizon simulation is conducted to evaluate the benefitsof integrating upgrading and revenue management under limited production capacity andcoproduction.

100

4.1 Introduction

In a limited capacity context, the decision of whether to accept or reject a customer ordercan affect the available capacity for future orders. Revenue management (RM) is a commonpractice which effectively allocates limited resources to more profitable demands in a profit-oriented organization (Hung, Tsai, and Wu 2014) : when expected demand exceeds the avai-lable capacity, sales managers have to select the most profitable orders. Consequently, orderscan be rejected in anticipation of more valuable ones (Guhlich, Fleischmann, and Stolletz2015), not only if not enough resources are available. Such an approach may be consideredunacceptable for some customers in manufacturing and may affect customer-manufacturerrelationships. Thus, in order to prevent losing sales, sales managers often propose other al-ternatives to fulfill customers orders such as offering substitutes or allowing partial orderfulfillment. They can also purchase from external sources to compensate their shortage.

This study focuses on the extension of the RM to practical applications in manufacturing. It ismotivated by the Canadian softwood lumber case where product substitution is a commondemand fulfillment practice, due to the co-production (i.e. from one log, a mix of high-valueand low-value products is generated). Canadian dimension lumber is sorted according tograding rules meeting Canadian and US requirements. Higher prices are attributed to higherquality products of the same dimensions. Based on the Canadian softwood lumber case, weconsider in this paper, first, that products of the same dimensions can be substitutable onlyif they belong to successive quality levels, which is called limited cascading. Second, it isassumed that softwood lumber customers always accept higher quality products if these areoffered at no extra cost. Thus, we consider only the situation when a higher quality substituteis provided at the original product’s price, which is called an upgrade. Upgrading can beparticularly beneficial if the selling firm faces stochastic and seasonal demand (Steinhardtand Gönsch 2012). In such a complex context, a simulation system can be used to evaluatethe benefits of integrating upgrading and RM concepts.

The paper contributions are i) to develop a generic consumption model integrating both RMand product substitution and ii) to evaluate, by means of simulation, the performance of anapproach integrating RM concepts and upgrading in a co-production context, compared tocommon demand fulfillment approaches. In addition to practical implications, the presentstudy also contributes to the existing demand fulfillment literature since we integrate dif-ferent consumption models, executed at the real-time level, with a Sales and OperationsPlanning (S&OP) model. Such tactical model additionally offers a medium-term visibility tomake supply and sales decisions. This integrated decision-support system is too complexto be evaluated analytically and has to be studied by means of simulation before imple-mentation. Thus, we use the platform previously developed by Ben Ali et al. (2014) for thesoftwood lumber industry, integrating S&OP at the tactical level and a consumption modelat the operational/execution level (see Figure 4.1), and we conduct a rolling horizon simula-

101

Context.png

FIGURE 4.1 – The integrated demand management process proposed by Ben Ali et al. (2014)

tion to compare different demand fulfillment approaches in different settings.

This paper is organized as follows : Section 4.2 provides a brief review of related literatureabout product substitution and RM in manufacturing contexts. In Section 4.3, we present thegeneric consumption model integrating both RM and product substitution and schematizethe interactions with the S&OP model. The proposed model can also be adapted to allowpartial order fulfillment and purchasing additional quantities. Section 4.4 describes the soft-wood lumber case and experiments that will be carried out. The results analysis is presentedin Section 4.5, followed by concluding remarks and future research opportunities in Section4.6.

4.2 Related literature

4.2.1 Product substitution

Product substitution offers flexibility for logistics and production systems. Among the exten-sive literature on product substitution, Lang (2010) provides an overview in the context ofproduction and inventory management, which aims to unify the conceptual framework andthe classification for product substitution models. The author particularly highlights thatsubstitution has been considered in some demand fulfillment publications, such as Chen,Zhao, and Ball (2001) and Fleischmann and Meyr (2004). Besides, in a more recent publica-tion, Ervolina et al. (2009) present an allocation model that comprises optimized availabili-ties of a firm’s core products as well as other product alternatives in an assemble-to-order(ATO) manufacturing environment, where end products are configured from standard com-ponents. Ervolina et al. (2009) assumed that supply quantities are exogenous inputs to themodel. However, in this paper, we use an integrated demand management process that notonly determines product allocations at short-term horizon, but also decides the ideal supplymix in terms of maximizing the firm’s profit by integrating allocation planning with a Salesand Operations Planning (S&OP).

4.2.2 Product substitution and revenue management

Many publications investigate RM with substitutable products in the service industry (seefor example, Petrick et al. (2010), Petrick et al. (2012), Steinhardt and Gönsch (2012) andGönsch, Koch, and Steinhardt (2014), as well as the books by Talluri and Van Ryzin (2004)

102

and Phillips (2005)). Products for which substitution is possible are often called flexible pro-ducts in RM (Zatta 2016). Another RM literature stream deals with substitution in manufac-turing context (Lang 2010). For instance, in Gurler, Oztop, and Sen (2009), Sibdari and Pyke(2010) and Kim and Bell (2011), prices are not given as data, but as decisions variables. Thesestudies consider substitution as a customer’s decision which occurs as a response to pricedifferences (i.e. if the price of a product increases, customers may look for less expensivealternatives), while in our paper product substitution is the firm’s policy and can affect theavailable capacity for future orders.

4.3 Mathematical formulation

In this paper, the S&OP model and different consumption models (see Figure 4.2) are for-mulated as linear programming (LP) models. Tables 4.1 and 4.2 present respectively sets,parameters, and decision variables involved in the consumption models and the S&OP allo-cation constraints.

TABLE 4.2 – Parameters and decision variables

Parameters DescriptionTimei First period of the mid-term horizon Indexj Current period IndexPrices/Costsαp,g,t Selling price of product p to segment g during period t $/Qtychol

m,p,t,t′Holding cost of product p in mill m from period t to period t

′$/Qty

FIGURE 4.2 – S&OP model and different consumption models

103

TABLE 4.2 – Parameters (continued)

Parameters Descriptioncpro

m,p Production cost of product p in mill m $/Qtyctra

m,p,g Transportation cost of product p from mill m to segment g $/Qtyβm,s,d Unit profit for selling quantities from mill m, initially set to s, to

demand d$/Qty

Consumption modelqd Total quantity required by d (Commitments + new order demand) Qtyxm,s Quantity from mill m allocated by the S&OP model to s (Decision

variables of S&OP model)Qty

ym,s,d Quantity from allocation xm,s already consumed QtyS&OP modelfs Demand forecast for s = (p, g, t), i.e. maximum demand to fulfill

of product p for segment g during period tQty

lbs Lower bound for quantities xm,s allocated to s Qtyubs Upper bound for quantities xm,s allocated to s Qtyinvm,p,t Inventories in mill m of product p at the end of period t QtyinvS&OP

m,p,t Inventories in mill m of product p at the end of period t plannedby the previous S&OP

Qty

Dec. variables DescriptionYm,s,d Quantity consumed from allocation xm,s to fulfill a demand d not

yet transported (t′ ≥ j)

Qty

104

TABLE 4.1 – Sets

Sets DescriptionM Mills mP Products pPSp’ Products p that can substitute product p

′(including p

′)

G Customer segments gT Periods t that are part of the short-term horizonD Set of d = (p

′, g′, t′) ∈ P×G× T (Demand)

S Set of s = (p, g, t) ∈ P×G× T (Consumption sources)Sd Set of s = (p, g, t) ∈ S considered as sources for a demand dSd Set of s = (p, g, t) ∈ S\Sd (Forbidden sources)Si Set of s = (p, g, t) ∈ S|t ≥ i (Consumption sources after period i)

4.3.1 Generic consumption model

We formulate a generic consumption model, which can be adapted for different demand ful-fillment approaches. This model has to be executed for each received order and is integratedto the S&OP model executed monthly, as presented in Figure 4.2.

We consider an assignment problem in which we can fulfill a demand requiring product p′

received from customer segment g′

for period t′

(i.e. a demand d = (p′, g′, t′)) from alloca-

tions of product p. Each allocation is initially set by the S&OP model to fulfill the demand ofa customer segment g for a period t(i.e. a source s = (p, g, t)). Nested booking limits are com-monly used (see for example Quante, Fleischmann, and Meyr (2009), Azevedo, D' Amours,and Rönnqvist (2016), Ben Ali et al. (2014)) in order to control the availability of capacity, sothat quantities initially allocated to s can only be sold to a demand d generating the same orbetter profits. We generalize the consumption model presented by Ben Ali et al. (2014) withnested booking limits, defined by customer segment and delivery period, and we add theproduct dimension to enable product substitution. We present also in Appendix G the ma-thematical formulation for two additional demand fulfillment approaches (allowing partialfulfillment and purchasing from external sources).

The goal of our consumption model is to maximize the short-term profit of fulfilling de-mand requested for periods between the current period j and the end of the short term hori-zon, which is expressed by Equation (4.1). This objective function is subject to the followingconstraints : First, constraints (4.2) ensure that quantities consumed from the quantity xm,s

available at mill m initially allocated to s = (p, g, t) will not exceed what is available. This in-cludes quantities ym,s,d consumed by delivered orders that we can no longer change (be reas-signed). Second, to guarantee previous commitments and new order fulfillment, constraints(4.3) are expressed so that order reassignment is allowed : quantities consumed to fulfill a

105

demand d have to be equal to the total quantity required by d, including the previous com-mitments and the quantity required by the new order. Otherwise, the new order cannot befulfilled. Third, constraints (4.4) force forbidden consumptions to be zero (see details below).Finally, constraints (4.5) ensure that all variables are non-negative.

Maximize ∑m∈M

∑d∈D

∑s∈Sd

βm,s,d Ym,s,d (4.1)

Allocation consumption

∑d∈D

(Ym,s,d + ym,s,d) ≤ xm,s ∀m ∈M, ∀s ∈ S (4.2)

Respect of previous commitments and fulfillment of the new order

∑m∈M

∑s∈Sd

Ym,s,d = qd ∀d ∈ D (4.3)

Forbidden consumptions

Ym,s,d = 0 ∀m ∈M, ∀d ∈ D, ∀s ∈ Sd (4.4)

Non-negativityYm,s,d ≥ 0 ∀m ∈M, ∀s ∈ S, ∀d ∈ D (4.5)

Unit profit calculation under upgrading hypothesis

In this study, we assume that higher selling prices are offered for higher quality productsand that the seller can fulfill a certain product request with a substitute from a pre-specifiedset of alternative products. With upgrading, the substitute is a higher quality product thanthe original product and the seller offers it at the requested product’s price. In other words,assuming that we receive an order from a customer from segment g

′requiring product p

′for

period t′, the customer accepts to substitute product p

′by a product p ∈ PSp

′at the price of

product p′. That is, the profit βm,s,d can be expressed by Equation (4.6).

βm,s,d = αp′ ,g′ ,t′ − cholm,p,t,t′ − cpro

m,p − ctram,p,g′ (4.6)

Forbidden consumptions

Naturally, to fulfill a demand d = (p′, g′, t′), some consumption decisions are not possible :

— With product substitution, we can substitute a required product p′

only by a few pro-ducts p ∈ PSp

′(PSp

′includes p

′). If product substitution is not allowed, PSp

′={

p′}

.

106

— We can consume from the allocations that are initially set to previous periods t (i.e.t < j) and that have not been consumed until current period j.

— In order to ensure the availability of the required quantity in the mill at the deliveryperiod t

′, we can only consume from allocations initially set to future periods t prece-

ding t′

(i.e. j ≤ t ≤ t′).

Furthermore with revenue management, we have to avoid consumptions to fulfill ordersfrom allocations set to more profitable demands. In order to consider the product substitu-tion, we add the product dimension to the expression of the nested booking limits definedby Ben Ali et al. (2014). Thus, to fulfill a demand d = (p

′, g′, t′) :

— we cannot consume from allocations xm,s initially set to s = (p, g′, t) where p ∈ PSp

′

and t = j..t′ − 1 (i.e. initially set to the same customer segment for different periods),

which can generate higher profit if sold to s rather than d.

— we cannot consume from allocations xm,s initially set to s = (p, g, t) where p ∈ PSp′,

g 6= g′

and t = j..t′

(i.e. initially set to different customer segments for any periodspreceding t

′), which can generate higher profit if sold to s rather than d.

To simplify, we formulate these restrictions in Equation (4.4) as we cannot consume fromallocations initially set to s = (p, g, t) ∈ Sd to fulfill a demand d = (p

′, g′, t′). Sd is expressed

differently depending on whether we consider or not revenue management and productsubstitution :

1st case : We do not consider RM and product substitution, i.e. we cannot consume fromallocations of products other than p

′.

Sd1 =

{s = (p, g, t) ∈ P\p

′ ×G× T, j < t < t′}

(4.7)

2nd case : We do not consider RM and we allow product substitution, i.e. we cannot consumefrom allocations of products p ∈ P\PSp

′

Sd2 =

{s = (p, g, t) ∈ P\PSp

′×G× T, j < t < t

′}

(4.8)

3rd case : We consider RM but not product substitution, i.e. we cannot consume from allo-cations of products other than p

′and we cannot consume from allocations which generate

higher profit if sold to s rather than d.

Sd3 =

{s = (p, g, t) ∈ P\p

′ ×G× T, j < t < t′}∪⋃

m∈M

{s = (p

′, g, t) ∈ G× T, j < t ≤ t

′and βm,s,s > βm,s,d

} (4.9)

4th case : We consider RM and product substitution i.e. we cannot consume from allocationsof products p ∈ P\PSp

′and we cannot consume from allocations which generate higher

107

profit if sold to s rather than d.

Sd4 =

{s = (p, g, t) ∈ P\PSp

′×G× T, j < t < t

′}∪⋃

m∈M

{s = (p, g, t) ∈ PSp

′×G× T, j < t ≤ t

′and βm,s,s > βm,s,d

} (4.10)

4.3.2 Interactions with the S&OP model

S&OP allocation constraints

The S&OP model includes supply, capacity, transport and flow balances constraints (Marieret al. 2014). In addition, allocation constraints have to consider demand forecasts and pre-vious commitments. Let us consider s = ( p, g, t) as the allocation sources of the S&OP modelexecuted at period i (i is the first period of the mid-term horizon and s ∈ Si ). The alloca-tions xm,s are decision variables in the S&OP model and are limited by lower lbs and upperbounds ubs (see Equation (4.11)). These bounds, expressed respectively in Equations (4.12)and (4.13), are considering the product substitution option :

— The lower bound lbs is the quantity committed to be consumed from an allocationsource s = ( p, g, t),

— The upper bound ubs is the maximum between the lower bound lbs and the forecast ofthe quantity that will be consumed from an allocation source s = ( p, g, t). This forecastis expressed in Equation (4.13) as the demand forecast f s, plus the quantity of p that willbe consumed to substitute other products p

′(p is a substitute of p

′), minus the quantity

that will be consumed to fulfill the demand of other products p (p is a substitute of p).

lbs ≤ ∑m∈M

xm,s ≤ ubs ∀s ∈ Si (4.11)

lbs = ∑m∈M

∑d|(p′ ,g,t)∈D

∑s|( p,g,t)∈Sd

ym,s,d ∀s ∈ Si (4.12)

ubs = max

lbs, f s + ∑m∈M

∑d|(p

′,g,t)∈D

p′ 6= p

∑s|( p,g,t)∈Sd

ym,s,d − ∑m∈M

∑d|( p,g,t)∈D

∑s|(p,g,t)∈Sd

p 6= p

ym,s,d

∀s ∈ Si (4.13)

108

End-of-period inventories

At the beginning of period i, the end-of-period inventories of period (i− 1) have to be trans-ferred to the S&OP model. As shown in Equation (4.14), the inventories in mill m at the endof period (i− 1) are computed as the inventories of period (i− 1) planned by the previousS&OP, plus the difference between what we plan to consume during period (i− 1) from millm by the previous S&OP (i.e. allocated quantities) and what we actually consume from millm (i.e. delivered quantities during period (i− 1)) .

invm, p,i−1 = invS&OPm, p,i−1 + ∑

s|(p,g,t)∈Sp= p

t=i−1

xm,s − ∑d|(p

′,g′,t′)∈D

t′=i−1

∑s|(p,g,t)∈Sd

p= p

ym,s,d

∀m ∈M, p ∈ P (4.14)

109

4.4 Application to softwood lumber case study

4.4.1 Case description

As presented by Figure 4.3 and Table 4.3, we consider a softwood lumber manufacturercomposed of three sawmills equipped with sawing, drying, and planing resources and loca-ted in Eastern Canada. In this region, the targeted markets are principally the NortheasternAmerican market (US), the Eastern Canadian market (CAE), and a spot market composed ofoccasional customers offering low prices.

FIGURE 4.3 – Supply network of a multi-site softwood company

TABLE 4.3 – Scope of the simulated case

Sets Size DetailsSawills 3 Sawing, drying and planing resources.2x4 family of products 6 10’ Grade 1&2, 10’ Grade 3, 10’ Grade 4

14’ Grade 1&2, 14’ Grade 3, 14’ Grade 4Markets 3 US, CAE and spot marketSegments 7 US and CAE markets are composed from 3 seg-

ments each. Spot market is considered as onesegment.

Average number oforders incoming weekly

100 Average weekly arrival rate is one order percouple (segment, product), where one productis required per order.

We consider six products from the 2x4 family. As a common practice in the Canadian soft-wood lumber industry, products having the same dimensions but different quality can sub-stitute each others. We use real market prices for the CAE market and the US market. Weassume that customers from the spot market, will offer low prices equivalent to 0.8 of the

110

US prices and that the manufacturer is able to purchase additional quantities from externalsources with a price equivalent to 0.9 of the local market price (CAE in our case).

In this paper, we consider substitution with limited cascading : substitution is allowed onlybetween products from successive quality levels. In addition, under the upgrading hypothe-sis, the substitutes of product from grade 3 are products from grade 1&2 and the substitutesof product from grade 4 are products from grade 3.

4.4.2 Demand generation and customer segmentation

The yearly global demand is considered as 150% or 200% of the maximum output that canbe produced by pushing an infinity of supply into the supply chain. Since the US marketrepresents the largest export market for Eastern Canadian softwood companies, we assumethat the demands of US market, CAE market and spot market represent respectively 50%,25% and 25% of the total demand.

Inspired from the real context, we assume that the spot market is considered as one segmentcomposed of occasional customers offering low prices (0.8 of the US market prices), whilethe US and CAE markets can be split in three customer segments each (see Table 4.4 ) :

— High-priority customers (10% of the market demand), typically home improvementwarehouse companies and housing component manufacturers, are ready to pay 20%more than the market price to have shorter transport lead times.

— Medium-priority customers representing the majority of customers (70% of the marketdemand) pay exactly the market price.

— Low-priority customers (20% of the market demand), typically dealers and distribu-tors, pay 20% less than the market price.

We assume that the manufacturer offers upgrades only to high-priority and medium-prioritycustomers.

TABLE 4.4 – Customer segments

Customer segments Demand Price

High-priority customers10% of the market de-mand

20% more than the mar-ket price

Medium-priority custo-mers

70% of the market de-mand

The market price

Low-priority customers20% of the market de-mand

20% less than the marketprice

Spot customers Spot market demand The lowest prices

111

4.4.3 Order generation

Randomness in our experiments concerns generating orders for the order promising leveland includes inter-arrival times, lead times and quantities required by customer orders. Wedeploy the same procedure described by Ben Ali et al. (2018) to generate different lists oforders for different ten replications. Assuming that we receive 100 orders weekly, i.e. 1 orderper combination (segment, product), we generate random variables as many as we have or-ders per combination (segment, product) in a year. For each order of a combination (segment,product) :

— We generate reception period based on inter-arrival times, which follow a Poisson dis-tribution. Average weekly arrival rate of order depends on product required and oncustomer segment.

— We generate delivery delays following a triangular distribution. Maximum, averageand maximum delays are respectively set to 1, 2 and 3 periods for high-priority seg-ments 1 and to 1, 3 and 4 periods for other segments. Then, we deduce delivery periods.

— We compute average quantity required by an order of a combination (segment, pro-duct) as weekly segment demand forecasts of the product demanded divided by theaverage weekly arrival rate. The quantity requested by an order is then deduced asthe inverse of a normal distribution using as mean the average value previously obtai-ned and 0.33×mean as a standard deviation such as Feng, D' Amours, and Beauregard(2008).

— We generate orders as a list ordered by reception date.

4.4.4 Simulation framework

The goal of the simulation is to highlight the benefits of integrating revenue management(RM) and upgrading (UPG). To this end, we simulate the behavior of the demand mana-gement process presented in Figure 4.2 with the consumption model RM-UPG, integratingRM concepts (i.e. using nested booking limits) and upgrading over a year. To evaluate thisintegrated approach, a comparison to common demand fulfillment approaches is required.Thus, we consider different consumption models as presented in Table 4.5 :

— The RM model is using nested booking limits to take into consideration customer hete-rogeneity and profitability variation over time. The RM-PF and RM-EXT models, usingalso nested booking limits, respectively allow partial fulfillment and purchasing fromexternal sources.

— The FCFS model is making consumption decisions based only on resource availabilityand without considering nested booking limits. FCFS-UPG, FCFS-PF and FCFS-EXTrespectively allow upgrading, partial fulfillment and purchasing from external sourceswithout considering nested booking limits.

112

TABLE 4.5 – Different consumption models

Model abbreviation Demand fulfillment approach

RM Revenue Management

RM-UPG RM + UPGrading

RM-PF RM + Partial Fulfillment

RM-EXT RM + EXTernal sources

FCFS First-Come First-Served

FCFS-UPG FCFS + UPGrading

FCFS-PF FCFS + Partial Fulfillment

FCFS-EXT FCFS + EXTernal sources

As an upper bound on the profit that we could get, we assumed an ideal process (oftencalled “oracle”), which knows all orders arriving within the planning horizon before makingpromises and enables a Global Optimization (GO) with a total visibility.

The S&OP model and the different consumption models presented in Table 4.5 are developedwithin IBM ILOG CPLEX Optimization Studio version 12.4. A rolling horizon simulation isconducted using an algorithm developed in Visual Basic. NET, which called sequentiallythe S&OP model (executed each month) and the consumption model (executed on order-by-order basis). Since we consider 100 orders/week, with 10 s for each order processing, a totalof 15 h is needed to perform a complete 1-year simulation. It should be noted that a warm-upperiod of 17 weeks is considered and that 10 replications were performed.

In what follows, we report the average values and the 95% confidence intervals of the yearlyprofit margin (to simplify we call it profit). Four different scenarios (see Table 4.6) are consi-dered, depending on the Demand Intensity (DI) relative to the production capacity and onthe Price Difference (PD) between successive quality levels.

A demand intensity DI equal to 100% is estimated as the maximum output that can be pro-duced by pushing an infinity of supply into the supply chain (considered as the capacity).Thus, we calculate the demand as : Demand = DI x Capacity. In the base case scenario, weconsider that DI is 150%. Similarly to Dumetz et al. (2015), we assume that the DI may reach200%. Regarding the PD, the base scenario case uses real prices. The price difference betweenproducts from grade 1&2 and grade 3 corresponds to 24% for 10’ products and 21% for 14’products. We can observe the same difference between products from grade 3 and grade 4.Then, we consider that this difference (i.e. the difference price between products from grade

113

TABLE 4.6 – Simulation scenarios

Scenarios Demand Intensity (DI)Price Difference between succes-sive quality levels (PD)

Scenario 1 ∗ 150%High (24% for 10’ products, 21%for 14’ products)

Scenario 2 150%Low (14% for 10’ products, 11%for 14’ products)

Scenario 3 200%High (24% for 10’ products, 21%for 14’ products)

Scenario 4 200%Low (14% for 10’ products, 11%for 14’ products)

∗ Base case scenario

1&2 and grade 3 and the difference price between products from grade 3 and grade 4) candecline until 10% smaller than the base case scenario. For this, we kept the same prices forthe products of grade 1&2 and we increased by 10% the prices of products of grades 3 and 4.

4.5 Results and discussion

In this section, we evaluate the performance of the approach integrating revenue manage-ment and upgrading compared to common demand fulfillment approaches in the base casescenario. Then, we investigate the effects of the sales price structure and the demand inten-sity on the performance of the integrated approach proposed.

4.5.1 Base case analysis

Figure 4.4 exhibits the profit generated by the different demand fulfillment approaches inthe base case scenario (i.e. the most realistic scenario for the softwood lumber context inEastern Canada). The value of integrating revenue management (RM) and upgrading (UPG)is presented in Figure 4.4 since the RM-UPG model achieves the highest profit (i.e. the closestto the upper bound on the profit that we could get with a Global Optimization GO).

Due to the limited capacity, approximately the same sales volume can be sold when pur-chasing from external resources is not allowed. In terms of profit, we can see that externalsources (RM-EXT and FCFS-EXT) approaches are less competitive. For this base case sce-nario, the RM-UPG and RM approaches achieve a profit gap which is well over 1 milliondollars compared to the approaches FCFS, FCFS-UPG, FCFS-PF, FCFS-EXT and RM-EXT.

In what follows, we will focus on upgrading approaches, which seem interesting (comparedto external sources and partial fulfillment approaches) in the most realistic scenario for thesoftwood lumber context in Eastern Canada.

114

FIGURE 4.4 – Profit of the consumption models in the base case scenario

4.5.2 Effect of the sales price structure and the demand intensity

Figure 4.5 illustrates the performance of the RM and FCFS approaches with and withoutupgrading in the four scenarios of Table 4.6. For each approach, we investigate how the profitvaries if the Demand Intensity (DI) or/and the Price Difference (PD) between successivequality levels change (for both DI and PD, we consider two levels as described in Table 4.6).In addition, the percentage of the profit achieved by each approach compared to the GO ispresented for each scenario in Figure 4.5.

Effect of the Demand Intensity (DI)

The difference between the FCFS and the FCFS-UPG with DI is not significant, due to thelimited capacity and the naive manner in which orders are promised. However, we can seethat the value of upgrading is significant if we integrated it with RM concepts for two rea-sons :

— At the S&OP level, each product is allocated to a specific customer segment. Since up-grading is offering a higher quality substitute at a lower price, a loss of profit is achie-ved if we provide the product to this specific segment or to other customer segmentspaying less than the original segment. This is particularly avoided by RM models bymeans of the nested booking limits.

— Avoiding unprofitable upgrades with RM preserves quantities for future profitable or-ders. Thus, RM-UPG is able to sell more than RM in all cases, and particularly for

115

(a) High price difference between successive quality levels

(b) Low price difference between successive quality levels

FIGURE 4.5 – Profit of the RM and FCFS approaches with/without upgrading in differentscenarios and the percentage of the profit achieved by each approach compared to the GO

high-priority customers. This explains the profit improvement for all cases in Figure4.5.

The gap between RM-UPG and RM decreases with the increasing demand. In fact, RM is ableto sell the products to customers who are asking for and then upgrades will not be necessaryto use up the inventories. We should note that the additional sales for RM and RM-UPG withDI of 200% are stabilizing above a certain level due to the limited production capacity.

Effect of the Price Difference between successive quality levels (PD)

As mentioned in Section 4.4.4, for low PD scenarios, we kept the same prices for the productsof grade 1&2 and we increased by 10% the prices of products of grades 3 and 4. This pricedifference is not significant to affect the performance of the simulated integrated demand

116

management processes (almost the same volumes are produced and sold with high PD andlow PD). This is due to the production recipes which are set to produce as much as possibleof high-value products (products of grade 1&2 in our case represent more than 75% of theproduction, which is common in sawmills).

4.6 Conclusion and future work

In this paper, we extend the research in demand fulfillment for co-production systems andinvestigate the benefits of integrating revenue management and upgrading compared tocommon demand fulfillment approaches.

First, we propose a linear programming model using nested booking limits and allowingproduct substitution. Based on the Canadian softwood lumber context, we assume that pro-ducts of the same dimensions can be substitutable only if they belong to successive qua-lity levels. Besides, we consider only upgrading, i.e. where a higher quality substitute isoffered at the original product’s price. Second, we evaluate alternative demand fulfillmentapproaches using rolling horizon simulation. Results in different scenarios enable the com-parison of multi-level systems composed of a Sales and Operations Planning (S&OP) modelat the tactical level coupled with different consumption models at real-time level.

Our simulation results demonstrate that integrating RM and upgrading achieves better per-formance than common demand fulfillment approaches in a context where demand exceedscapacity. The value of upgrading is more significant when integrated with RM concepts,since the use of nested booking limits prevents from doing unprofitable upgrades. Thus, in-ventories are preserved for future profitable orders. In the softwood lumber industry, thevalue of upgrading is not sensitive to the price difference between products from successivequality levels since high-value products represent the majority of production. The benefitsof upgrading are, however, less significant when the demand intensity increases.

Future works will be to analyze other substitution policies, such as downselling and up-selling, which can be valid in industrial contexts other than the softwood lumber industry.Considering the sensitivity of customers to substitution may also be of theoretical and prac-tical interest.

117

Appendix G

Additional demand fulfillmentapproaches

We extend the generic consumption model presented in Section 4.3.1 to allow partial ful-fillment and purchasing additional quantities from external sources. Table G.1 describes ad-ditional parameters and decision variables used.

TABLE G.1 – Additionnal parameters and decision variables

Parameters DescriptionPrices/Costscext

p,t Purchasing cost of product p at period t from externalsources

$/Qty

βextm,d Profit for consuming a unit purchased from external

sources and transported from mill m to fulfill a demandd

$/Qty

Quantitiesqcom

d Commitments to d already made in previous consump-tion cycles

$/Qty

λ Proportion of the quantity required by a new order thathave to be satisfied with partial fulfillment

Qty

Dec. variables DescriptionEm,d Quantity purchased from external sources and trans-

ported from mill m to fulfill a demand dQty

G.0.1 Allowing partial fulfillment :

If partial fulfillment is allowed, Equation (4.3) should be replaced by Equations (G.1) and(G.2). Constraints (G.1) ensure that the quantity fulfilled for d is more than what was com-mitted before and less than what is required for d including the new order. Constraints G.2force that a minimal proportion λ of the quantity required by a new order have to be sa-tisfied. Otherwise, the new order cannot be fulfilled. If the new order is accepted, the new

118

commitment qcomd for the demand d will be ∑m∈M ∑s∈Sd Ym,s,d − qcom

d .

Respect of previous commitments and partial fulfillment of the new order

qcomd ≤ ∑

m∈M∑

d∈DYm,s,d ≤ qd ∀d ∈ D (G.1)

λ× (qd − qcomd ) ≤ ∑

m∈M∑

s∈Sd

Ym,s,d − qcomd ∀d ∈ D (G.2)

G.0.2 Purchasing from external sources :

The unit profit βextm,d for purchasing a unit from external sources and selling it from mill m to

fulfill d = (p′, g′, t′) is expressed by Equation (G.3) as selling price minus purchasing cost

at period t′

and transport cost from mill m to customer segment g′. We assume that there is

always external sources near mills, thus transport delays between them are not considered.

βextm,d = αp′ ,g′ ,t′ − cext

p′ ,t′ − ctram,p′ ,g′ (G.3)

Thus, the objective function expressed by Equation (4.1) of the generic model should bereplaced by Equation (G.4), so that we can consider the external purchase cost included inthe unit profit βext

m,d expressed by Equation (G.3). In order to accept purchasing only if we havea positive unit profit, Equation (G.5) is added. Moreover, Equation (4.3) of the generic modelis replaced by Equation (G.6) to consider the quantities purchased from external sources.

Maximize ∑m∈M

∑d∈D

∑s∈Sd

βm,s,d Ym,s,d + ∑m∈M

∑d∈D

βextm,dEm,d (G.4)

Positive profitβext

m,dEm,d ≥ 0 ∀m ∈M, ∀d ∈ D (G.5)

Respect of previous commitments and fulfillment of the new order

∑m∈M

∑s∈Sd

Ym,s,d + ∑m∈M

Em,d = qd ∀d ∈ D (G.6)

119

Conclusion générale et perspectives derecherche

5.1 Conclusion générale

Tout au long de cette thèse, nous nous sommes intéressés à la problématique de la gestion dela demande dans l’industrie du bois d’œuvre. Les défis économiques, ainsi que la conjonc-ture difficile affectant le secteur, exigent de développer des systèmes intégrés d’aide à ladécision. À cette fin, nous avons combiné des outils de recherche opérationnelle et de simu-lation, soit la programmation linéaire et la simulation en horizon roulant, avec des conceptsvariés de planification et d’allocation de la capacité à différents niveaux décisionnels. Enoutre, nous avons utilisé des procédures novatrices d’analyse et considéré plusieurs scéna-rios afin de valider nos résultats.

Les contributions scientifiques de cette thèse portent sur la valeur de l’intégration des sys-tèmes de gestion de la demande dans une perspective de maximisation des revenus. Afin demener ces travaux, nous avons dû réunir trois domaines de recherche qui ont été rarementliés dans la littérature, soit la planification des ventes et des opérations (S&OP), la gestion desrevenus (Revenue management RM) et les chaînes de valeur dans l’industrie des produitsforestiers. Ces travaux relèvent également des défis importants, outre le développement demodèles d’optimisation et d’une plateforme de type simulation - optimisation captant lesrétroactions entre différents niveaux de décision. En effet, des efforts supplémentaires ontété dédiés à la collecte des données d’une chaîne d’approvisionnement du bois d’œuvre (ap-provisionnement, production, transport, stockage, demande ...), au choix du cas d’étude età la définition des modèles à comparer. De plus, vu le nombre des expériences à réaliserpour plusieurs modèles et avec des paramètres variés, il a été primordial de développer unoutil de génération des données de demande et des prix et d’automatiser la récupération etl’analyse des résultats.

L’utilisation de la simulation est très répondue dans la recherche en gestion des opérations(Law 2015). En particulier dans cette thèse, la simulation a été le principal outil utilisé pourreproduire le comportement d’un processus réel de ventes et d’opérations dans le but d’éva-

120

luer différents processus et approches intégrés. Traiter les commandes d’une année complètede façon séquentielle, tel est le cas dans le monde réel, a fait qu’on a des expériences cou-teuses en terme de temps à simuler ( dans le cas où on reçoit 100 commandes par semaineet que le traitement d’une commande dure 10 s, ce qui est très acceptable dans le monderéel, une expérience complète nécessite 100 commandes/semaine x 52 semaines x 10s, soitun total de 15h). Il a été donc nécessaire de réaliser des simulations en parallèle (avec unemachine dotée de 10 processeurs) et de définir judicieusement le plan d’expériences à suivrepour chaque contribution. Naturellement, il a fallu réaliser un ensemble d’essais-erreurs àchaque fois avant d’arriver au plan d’expériences final.

Cette thèse apporte de nouvelles approches intégrées pour supporter la prise de décisiondes ventes dans l’industrie du bois d’œuvre. Les résultats obtenus, bien que leur validité soitlimitée aux cas considérés, amènent de nouvelles réflexions sur des pratiques prometteusespour le secteur et peuvent être adaptés à des études dans d’autres secteurs industriels.

Dans ce qui suit, nous présenterons une synthèse des différentes contributions de cette thèse,ainsi que les limites de ces travaux de recherche.

5.1.1 Intégrer la gestion des revenus et la planification des ventes et desopérations dans un environnement de fabrication pour les stocks

Dans notre première contribution, nous avons présenté un processus intégrant deux ap-proches communes de gestion de la demande, soit la planification des ventes et des opé-rations (S&OP) et la gestion des revenus (Revenue management RM). Un cadre décisionnelmultiniveau a été proposé compte tenu les meilleurs pratiques en gestion de la demandedans la littérature, ainsi que les systèmes de technologies d’information utilisés par les par-tenaires du consortium FORAC. Une plateforme d’optimisation et de simulation en horizonroulant, capturant les rétroactions entre les différents niveaux de décisions, a été développéeafin de soutenir les praticiens, ayant une expérience limitée avec le S&OP et le RM, face auxdéfis d’implantation.

La performance des systèmes intégrés d’aide à la décision dans un contexte manufacturierest généralement difficile à mesurer analytiquement. Nous avons donc proposé d’évaluer,via la simulation, i) les bénéfices d’implanter un processus S&OP, ii) les bénéfices d’utiliserles limites de réservation imbriquées, et iii) les bénéfices d’implanter un processus intégrantle S&OP et les limites de réservation imbriquées. Les résultats de simulation ont démontré lacapacité d’un processus intégrant le S&OP et le RM, dans un contexte de capacité limitée etface à une clientèle hétérogène, à réaliser des meilleurs profits et à mieux satisfaire les clientsprioritaires que les processus conventionnels de gestion de la demande.

La plateforme d’optimisation et de simulation développée est un outil pouvant être utilisépar des praticiens de l’industrie du bois d’œuvre dans le but de simuler de nouveaux mo-

121

dèles d’affaires et d’anticiper la performance d’un processus intégrant le S&OP et le RMdans différentes situations, telles que l’introduction de nouveau produit, le changement dela capacité de production, la pénétration d’un nouveau marché, etc.

5.1.2 Évaluer la performance de différentes approches intégrées de gestion de lademande face à une variété de scénarios du marché

Lors de la deuxième contribution, nous avons utilisé la plateforme développée dans la pre-mière contribution afin de simuler différentes configurations du processus intégré face àune variété de scénarios du marché, ceci en considérant trois séquences d’arrivée des com-mandes. En outre, nous avons proposé une procédure novatrice d’analyse des résultats :nous avons utilisé un plan de remplissage d’espace pour identifier efficacement les scéna-rios à simuler et des métamodèles de krigeage pour évaluer l’effet des facteurs environne-mentaux analysés (soit, l’intensité de la demande comparée à la capacité, la précision desprévisions, l’hétérogénéité des clients manifestée par la différence entre les prix offerts parles multiples segments de clients, la variabilité de la taille des commandes) sur la perfor-mance de processus intégrés de gestion de la demande.

Les résultats de la simulation ont démontré l’impact de l’approche de promesse de livrai-son, ainsi que de la séquence d’arrivée des commandes sur le profit annuel et le niveau desatisfaction des clients prioritaires. En particulier, l’étude a mis en évidence l’améliorationpotentielle de performance qu’on peut atteindre en utilisant les concepts de RM. L’analysedes métamodèles de krigeage, illustrant l’impact des facteurs environnementaux sur la per-formance du processus intégré de gestion de la demande, a démontré que l’intensité de lademande et la précision des prévisions sont deux facteurs pertinents qui peuvent affecter laperformance de l’entreprise. Outre l’analyse via différents outils statistiques et visuels, unedescription détaillée des implications managériales de cette analyse a été présentée.

Cette étude a des implications pour les chercheurs en gestion de la demande, ainsi que pourles praticiens. Notre contribution pour la recherche est de mettre en évidence la valeur d’in-tégrer deux concepts communs en gestion de la demande, soit le S&OP et le RM, ainsi qued’utiliser un plan de remplissage d’espace et la technique de krigeage dans un contexte dechaîne d’approvisionnement. On offre également aux praticiens un outil d’aide à la déci-sion, ainsi qu’une procédure efficace pour évaluer un processus complexe dans un contextede gestion de la chaîne d’approvisionnement.

5.1.3 Simuler une approche intégrée de gestion des revenus pour un système decoproduction avec substitution de produits

Dans la troisième contribution, nous avons investigué l’intérêt d’intégrer la substitution desproduits et les concepts de RM dans un contexte de capacité limitée. Particulièrement, nous

122

nous sommes intéressés à l’effet de proposer à certains clients privilégiés un produit de qua-lité supérieure au prix du produit original demandé, soit l’équivalent d’un sur-classement("upgrading") pour les entreprises de service.

Afin d’introduire la substitution des produits, nous avons développé un modèle génériquequi peut être adapté pour différentes approches de promesse de livraison et nous avonsutilisé la plateforme développée dans la première contribution pour simuler leurs perfor-mances dans un contexte de capacité limitée. En partant du scénario le plus réaliste pour uneindustrie québecoise du bois d’œuvre, nous avons comparé plusieurs approches intégrées,soit l’application du RM en utilisant les limites de réservation imbriquées, le sur-classement,la satisfaction partielle des commandes et l’approvisionnement à partir de sources externes.Ensuite, nous avons évalué la valeur de l’intégration des concepts de RM et du sur-classementface à différents scénarios du marché.

Cette étude souligne l’efficacité d’une approche de promesse de livraison intégrant le sur-classement et les concepts de RM, comparée aux pratiques communes adoptées par les par-tenaires industriels du consortium FORAC afin de satisfaire la demande dans le contextede capacité limitée. Tout d’abord, les résultats de la simulation ont confirmé qu’utiliser leslimites de réservation imbriquées permet de réaliser de meilleurs profits que de permettrela satisfaction partielle des commandes ("partial fulfillment"). Les résultats ont égalementdémontré que combler les pénuries en s’approvisionnant à partir de ressources externes,même à un coût raisonnable (dans notre cas, -10% que le prix de vente local), n’est pas uneapproche compétitive en terme de profit, bien qu’elle permet de réaliser plus de ventes. Parcontre, le sur-classement (càd, proposer un substituant de qualité supérieure au prix du pro-duit original demandé) s’avère significativement bénéfique s’il est utilisé judicieusement,ceci en considérant les concepts de RM. En effet, l’utilisation de limites de réservation em-pêche de proposer un sur-classement si le produit en question a été alloué à des commandesplus rentables (càd, pour des clients plus payants et/ou des périodes plus payantes).

Le modèle de promesse de livraison proposé, intégrant la substitution des produits et lesconcepts de RM, pourrait être utilisé afin d’investiguer l’intérêt d’adopter des politiques desubstitution dans un contexte manufacturier autres que le sur-classement, soit les stratégiesde montée en gamme 1 ou de baisse en gamme 2. Celles-ci peuvent être applicables dans desindustries autres que l’industrie du bois d’œuvre (Ervolina et al. 2009).

5.1.4 Limites

Tout le long de cette thèse, nous avons adopté certaines hypothèses :

1. Technique de vente qui consiste à inciter le consommateur à acheter un produit (un substituant) plus cherque celui qu’il avait d’abord choisi, en lui en proposant un d’une gamme supérieure.

2. Technique de vente qui consiste à inciter le consommateur à acheter un produit (un substituant) moinscher que celui qu’il avait d’abord choisi, en lui en proposant un d’une gamme inférieure.

123

Premièrement, nous avons considéré un mode de production pour les stocks. Or pour lesentreprises du bois d’œuvre, d’autres modes de production pourraient être mis en place :certains produits spécifiques, généralement plus dispendieux que les produits de commo-dité, peuvent être fabriqués sur commande.

Deuxièmement, les décisions d’approvisionnement, de production et de transport ont étéprises seulement au niveau tactique, en supposant qu’elles seront réalisables au niveau opé-rationnel. En réalité, les plans opérationnels d’approvisionnement, de production et de trans-port doivent être pris en considération.

Troisièmement, nous avons supposé dans cette thèse que les prix réels de vente corres-pondent exactement aux prix estimés par prévision. De plus, les prix de vente ont été consi-dérés comme des données exogènes. Or, la fixation des prix, contribuant en grande partie àla maximisation des revenus, comprend un ensemble de décisions qui doivent prendre enconsidération trois principaux facteurs soit, les coûts, la concurrence et les clients.

Finalement, peu d’attention a été accordée aux autres composantes du réseau d’approvi-sionnement. Particulièrement, cette thèse ne tient compte ni de l’effet de la compétition surle processus du vente, ni du comportement des clients en cas de ventes perdues.

5.2 Perspectives de recherche

Les perspectives de recherche en gestion de la demande se concentrent sur l’avenir d’une ap-proche davantage centrée sur le client. L’étude récente de Currie et al. (2018) prévoit que lesorientations futures de la recherche en gestion de la demande auraient une liaison étroiteavec des enjeux d’analyse des données massives, soit principalement : la valorisation etl’échange des données sur les clients, ainsi que l’analyse des comportements des clients enprésence de multitude de produits et tarifs. Dans ce contexte, on peut citer comme pistes derecherche pertinentes :

5.2.1 Développer des méthodes intelligentes d’établissement de prix ("pricing")

Les prix figurent parmi les décisions importantes à prendre par les responsables des ventes.McKinsey & Company consulting 3 estime que 30% les décisions des prix ne sont pas opti-males (Baker, Kiewell, and Winkler 2009), ce qui correspond à un grand manque à gagner.Vu le nombre de clients et des canaux de distribution qui ne cessent d’augmenter en expo-nentiel, il est primordial de saisir les opportunités offertes par les données massives pourdéterminer les meilleurs prix pour chaque produit et à chaque période.

3. McKinsey & Company consulting est un cabinet de conseil classé à la première position du classementdes cinquante meilleurs cabinets de consultation mondiaux. En 2002, McKinsey conseille 147 des 200 premièresentreprises mondiales. Voir https://www.mckinsey.com/ca/fr

124

En particulier pour les entreprises du bois d’œuvre, les responsables doivent identifier, pourchaque produit, le prix que le client acceptera de payer. À cette fin, ils doivent analyserpériodiquement différents facteurs qui peuvent influencer le client, voir le prix des produitsconcurrents et la valeur du produit chez le client. Cette tâche demande beaucoup d’effortd’autant plus que le nombre de produits augmente, ce qui correspond à un défi de donnéesmassives.

Afin de trouver le prix optimal, il faut utiliser des systèmes automatisés pour l’analyse desdonnées de prix (par exemple, déterminer ce qui a de la valeur pour chaque client en ana-lysant l’historique de transactions). Des clients de McKinsey & Company consulting parexemple, provenant de différentes industries (industrie chimique, construction, télécommu-nications. . . ), ont réussi à réaliser des marges de profit allant jusqu’à 20% en analysant lesdonnées massives pour la prise des décisions de prix (Baker, Kiewell, and Winkler 2009). Parconséquent, il est raisonnable de penser qu’un grand potentiel de gain peut être atteint enutilisant des systèmes automatisés pour l’analyse des données de prix du bois d’œuvre.

5.2.2 Instaurer des méthodes novatrices de prévision de la demande et du prix

Dans les trois contributions de cette thèse, on a supposé que les prévisions de la demandeet des prix sont fournies. En réalité, les prévisions peuvent être générées par des techniquesvariées et à partir d’un processus complexe qui peut impliquer différentes entités de la chaîned’approvisionnement.

État des lieux dans l’industrie du bois d’œuvre du Québec

Généralement, les entreprises du bois d’œuvre du Québec ne conservent que l’historique desventes réalisées. De plus, si un substituant a été offert, la demande originale n’apparaît pasdans les bases de données (Lemieux et al. 2008). Or, afin de fournir des bonnes prévisions, ilest nécessaire de conserver l’historique complet de la demande reçue et non seulement desventes réalisées. Actuellement, pour les entreprises du bois d’œuvre du Québec, les prévi-sions de la demande sont établies en se basant sur l’historique des ventes réalisées en utili-sant des méthodes ARIMA (Autoregressive Integrated Moving Average). Les prévisions desprix, quant à eux, sont produits de façon centralisée par le Conseil de l’industrie forestièredu Québec (CIFQ). Cependant, des méthodes qualitatives basées sur le jugement sont utili-sées hebdomadairement, ce qui nécessite un grand effort d’analyse manuelle vu le volumede données traitées.

Pistes de recherche pour les entreprises du bois d’œuvre du Québec

Parmi les pistes de recherche dans l’industrie du bois d’œuvre, il serait intéressant de dé-montrer l’intérêt d’un processus de prévision collaboratif ("collaborative forecasting"). Plu-sieurs études récentes (voir par exemple Yao et al. (2013), Eksoz, Mansouri, and Bourlakis

125

(2014), Gao (2015), Galbreth, Kurtulus, and Shor (2015)) ont démontré que la précision desprévisions est améliorée lorsqu’on passe d’un scénario non collaboratif (où le détaillant etle fournisseur produisent leurs prévisions indépendamment sans aucun partage d’informa-tion) à un scénario collaboratif (où les prévisions sont produites de manière centralisée).

Afin d’établir les prévisions de de la demande et des prix de façon plus efficace, plusieurspistes d’amélioration peuvent être explorées. Comme première étape, il faut commencer paranalyser les bonnes données, soit l’historique de la demande reçue et non des ventes réa-lisées. Ensuite, il est possible de tester de nouvelles méthodes de prévision de la demandeet des prix autres que les méthodes ARIMA et les méthodes conventionnelles basées sur lejugement.

L’étude récente de Torbat, Khashei, and Bijari (2018) sur les prévisions pour les produits decommodités a pointé les limites des méthodes ARIMA (principalement l’hypothèse de linéa-rité entre les valeurs futures prévues et les valeurs passées de la série temporelle) et a démon-tré qu’on peut réaliser des meilleurs prévisions en utilisant des modèles hybrides (appelés"hybrid probabilistic fuzzy ARIMA models"). Il s’agit d’une combinaison d’outils d’intel-ligence computationnelle ("computational intelligence") et de technologies programmables("soft computing techniques"). En outre, plusieurs recherches (voir par exemple, Schneiderand Gupta (2016), Fang, Jiang, and Song (2016), Papanagnou and Matthews-Amune (2017),Villegas, Pedregal, and Trapero (2018), Murray, Agard, and Barajas (2018)) ont proposé desoutils d’analyse de données massives afin d’atteindre des prévisions plus précises.

5.2.3 Mettre en œuvre une approche de gestion de la demande et de gestion desrevenus (Revenue Management RM) intégrant un système de gestion desrelations clients (Customer Relationship Management CRM)

État des lieux

Bien que les systèmes de CRM et de RM soient considérés complémentaires, l’intégration desdeux concepts a été peu abordée (Vaeztehrani, Modarres, and Aref 2015), en particulier dansun secteur autre que le secteur de service. La littérature existante accorde peu d’importanceà l’effet long terme de l’application des pratiques de RM (par exemple, le fait de donner plusde priorité aux clients peu sensibles aux prix peut être mal perçu par les autres clients) endésaccord avec l’objectif principal d’une approche CRM, soit préserver des relations à longterme avec les clients.

L’analyse des données massives appliquée au CRM

Un système CRM permettrait la collecte de renseignements utiles pour déterminer les be-soins actuels et potentiels des clients et analyser leurs comportements. Or, le grand volumed’information qui peut être généré par ces systèmes doit être exploité. On a donc besoin de

126

traiter et d’analyser ces données pour qu’elles aboutissent à une création de valeur pour lesentreprises. Une des utilisations potentielles de tels types de données serait de les intégrerdans un système de gestion des revenus et de les utiliser pour améliorer les prévisions de lademande afin de promouvoir au bon moment, au bon client, au bon prix et par le bon médiace qu’on offre comme produits. Une entreprise qui maîtrise la compréhension et l’interpré-tation de ces données aura un véritable avantage concurrentiel à gagner.

Défis de l’analyse des données massives pour un système intégrant le RM et le CRM

Un premier défi de l’application de l’analyse des données massives au CRM serait de tenircompte de la législation sur la protection des données. Un deuxième défi serait la collecteet l’organisation des renseignements nécessaires, ainsi que l’incorporation des données re-cueillies de façon instantanée afin d’exploiter ces informations par la suite à des fins de pla-nification. Les logiciels et les systèmes utilisés aujourd’hui dans les entreprises ne sont pasadaptés à l’exploitation des données massives vu qu’ils n’ont pas été choisis pour atteindrecet objectif, mais plutôt pour des fins de gestion des opérations et de logistique.

En outre, la mise en œuvre d’un système de RM & CRM peut s’accompagner de craintesde la part des entreprises manufacturières, qui redouteraient de l’efficacité d’une approcheintégrée RM & CRM à la fois en termes d’efficacité et de rapidité. Il est important d’évaluerpar simulation la performance d’une telle approche intégrée dans différents scénarios dumarché en utilisant des modèles stochastiques et des modèles d’analyse du risque.

127

Bibliographie

Affonso R, Marcotte F, Grabot B, 2008 Sales and operations planning : The supply chain pillar.Production Planning and Control 19 :132–141.

Al-Ameri TA, Shah N, Papageorgiou LG, 2008 Optimization of vendor-managed inventory sys-tems in a rolling horizon framework. Computers and Industrial Engineering 54(4) :1019 – 1047.

APICS, ed., 2013 APICS Dictionary. 14 edition, ISBN 978-0-9882146-1-3.

Assarzadegan P, Rasti-Barzoki M, 2016 Minimizing sum of the due date assignment costs, maxi-mum tardiness and distribution costs in a supply chain scheduling problem. Applied Soft Compu-ting 47 :343 – 56.

Azadeh A, Zarrin M, Salehi N, 2016 Supplier selection in closed loop supply chain by an integratedsimulation-Taguchi-DEA approach. Journal of Enterprise Information Management 29(3) :302 –26.

Azevedo RC, D' Amours S, Rönnqvist M, 2016 Advances in profit-driven order promising forMake-To-Stock environments – a case study with a canadian softwood lumber manufacturer. IN-FOR : Information Systems and Operational Research 54(3) :210–233.

Baker W, Kiewell D, Winkler G, 2009 Using big data to make better pricing de-cisions. URL https://www.mckinsey.com/business-functions/marketing-and-sales/

our-insights/using-big-data-to-make-better-pricing-decisions, accessed : 2018-05-25.

Bandaly D, Satir A, Shanker L, 2014 Integrated supply chain risk management via operationalmethods and financial instruments. International Journal of Production Research 52(7) :2007 – 25.

Ben Ali M, D' Amours S, Gaudreault J, Carle MA, 2018 Configuration and evaluation of anintegrated demand management process using a space-filling design and kriging metamodeling.Operations Research Perspectives 5 :45 – 58.

Ben Ali M, Gaudreault J, D' Amours S, Carle MA, 2014 A multi-level framework for demandfulfillment in a Make-To-Stock- A case study in Canadian softwood lumber industry. MOSIM2014 (Nancy, France).

128

Bottani E, Montanari R, 2010 Supply chain design and cost analysis through simulation. Interna-tional Journal of Production Research 48(10) :2859 – 86.

Boulaksil Y, 2016 Safety stock placement in supply chains with demand forecast updates. OperationsResearch Perspectives 3 :27 – 31.

Cavazzuti M, 2013 Optimization methods from theory to design scientific and technological aspectsin mechanics (Berlin : Springer).

Chen CY, Zhao ZY, Ball M, 2001 Quantity and due date quoting available to promise. InformationSystems Frontiers 3(4) :477–488.

Chen-Ritzo CH, Ervolina T, Harrison TP, Gupta B, 2010 Sales and operations planning in systemswith order configuration uncertainty. European Journal of Operational Research 205 :604–614.

Chevalier P, Lamas A, Lu L, Mlinar T, 2015 Revenue management for operations with urgentorders. European Journal of Operational Research 240(2) :476 – 487.

Chiang DMH, Wu AWD, 2011 Discrete-order admission ATP model with joint effect of margin andorder size in a MTO environment. International Journal of Production Economics 133(2) :761 –775.

Ciancimino E, Cannella S, Bruccoleri M, Framinan JM, 2012 On the bullwhip avoidance phase :The synchronised supply chain. European Journal of Operational Research 221(1) :49 – 63.

Coleman DE, Montgomery DC, 1993 A systematic approach to planning for a designed industrialexperiment. Technometrics 35(1) :1–12.

Crum C, Palmatier G, 2003 Demand Management Best Practices : Process, Principles and Collabo-ration (J. Ross Publishing), 1st edition edition, ISBN 9781932159011 AND 9781932159011.

Currie CSM, Dokka T, Harvey J, Strauss AK, 2018 Future research directions in demand mana-gement. Journal of Revenue and Pricing Management .

Dansereau LP, El-Halwagi M, Mansoornejad B, Stuart P, 2014 Framework for margins-basedplanning : Forest biorefinery case study. Computers and Chemical Engineering 63 :34 – 50.

Dev N, Shankar R, Debnath R, 2014 Supply chain efficiency : a simulation cum DEA approach.International Journal of Advanced Manufacturing Technology 72(9-12) :1537 – 49.

Dominguez R, Cannella S, Framinan J, 2015 The impact of the supply chain structure on bullwhipeffect. Applied Mathematical Modelling 39(23) :7309 – 25.

Duarte Canever M, Van Trijp HC, Beers G, 2008 The emergent demand chain management : keyfeatures and illustration from the beef business. Supply Chain Management : An InternationalJournal 13(2) :104–115.

129

Dufour D, 2007 L'industrie canadienne du bois d'oeuvre :tendances récentes. Statistiques Canada3–16.

Dumetz L, Gaudreault J, Thomas A, Marier P, Lehoux N, el Haouzi H, 2015 A simulationframework for the evaluation of production planning and order management strategies in the saw-milling industry. IFAC-PapersOnLine 48(3) :622 – 627, iNCOM 2015.

Eksoz C, Mansouri SA, Bourlakis M, 2014 Collaborative forecasting in the food supply chain : Aconceptual framework. International Journal of Production Economics 158 :120 – 135.

Ervolina T, Ettl M, Lee Y, Peters D, 2009 Managing product availability in an assemble-to-ordersupply chain with multiple customer segments. OR Spectrum 31(1) :257–280.

Fang K, Jiang Y, Song M, 2016 Customer profitability forecasting using big data analytics : A casestudy of the insurance industry. Computers and Industrial Engineering 101 :554 – 564.

Favreau J, Ristea C, 2017 The roles of FPInnovations, governments and industry in transformingCanada's forest industry. D' Amours S, Ouhimmou M, Audy JF, Feng F, eds., Forest valuechain optimization and sustainability, 91 (Boca Raton, FL : CRC Press), ISBN 9781498704861.

Feng Y, D' Amours S, Beauregard R, 2008 The value of sales and operations planning in orientedstrand board industry with Make-To-Order manufacturing system : Cross functional integrationunder deterministic demand and spot market recourse. International Journal of Production Econo-mics 115 :189–209.

Feng Y, D' Amours S, Beauregard R, 2010 Simulation and performance evaluation of partiallyand fully integrated sales and operations planning. International Journal of Production Research48 :5859–5883.

Feng Y, Martel A, D' Amours S, Beauregard R, 2013 Coordinated contract decisions in a Make-To-Order manufacturing supply chain : A stochastic programming approach. Production and Ope-rations Management 22(3) :642–660.

Feng Y, Xiao B, 2000 Revenue management with two market segments and reserved capacity forpriority customers. Advances in Applied Probability 32(3) :800–823.

Fleischmann B, Meyr H, 2003 Planning Hierarchy, Modeling and Advanced Planning Systems,volume 11, 455–523 (Elsevier).

Fleischmann B, Meyr H, 2004 Customer orientation in advanced planning systems. Dyckhoff H,Lackes R, Reese J, eds., Supply chain management and reverse logistics, xviii, 426 p. (Berlin :Springer), ISBN 3540404910.

Forget P, D' Amours S, Frayret JM, Gaudreault J, 2009 Study of the performance of multi-behaviour agents for supply chain planning. Computers in Industry 60(9) :698 – 708.

130

Frayret JM, D’Amours S, Rousseau A, Harvey S, Gaudreault J, 2007 Agent-based supply-chainplanning in the forest products industry. International Journal of Flexible Manufacturing Systems19(4) :358–391.

Frohlich MT, Westbrook R, 2002 Demand chain management in manufacturing and services : web-based integration, drivers and performance. Journal of Operations Management 20(6) :729–745.

Galbreth MR, Kurtulus M, Shor M, 2015 How collaborative forecasting can reduce forecast accu-racy. Operations Research Letters 43(4) :349 – 353.

Gao L, 2015 Collaborative forecasting, inventory hedging and contract coordination in dynamic sup-ply risk management. European Journal of Operational Research 245(1) :133 – 145.

Gao L, Xu SH, Ball MO, 2012 Managing an Available-To-Promise assembly system with dynamicshort-term pseudo-order forecast. Management Science 58(4) :770–790.

Gaston C, Robichaud F, 2017 Overview of wood product markets, distribution, and market re-search in North America. implications for value chain optimization. D' Amours S, OuhimmouM, Audy JF, Feng F, eds., Forest value chain optimization and sustainability, 181 (Boca Raton,FL : CRC Press), ISBN 9781498704861.

Gaudreault J, Forget P, Frayret JM, Rousseau A, D’Amours S, 2010 Distributed operations plan-ning in the softwood lumber supply chain : Models and coordination 17.

Gönsch J, Koch S, Steinhardt C, 2014 Revenue management with flexible products : The value offlexibility and its incorporation into DLP-based approaches. International Journal of ProductionEconomics 153 :280 – 294.

Goupy J, 2015 Plans d’expériences. URL https://www.techniques-ingenieur.fr/

base-documentaire/mesures-analyses-th1/qualite-au-laboratoire-42497210/

plans-d-experiences-p230/, accessed : 2018-05-25.

Government of Quebec, 2016 Competitiveness in the Quebec forest industry : budget 2016-2017.Technical report.

Guhlich H, Fleischmann M, Stolletz R, 2015 Revenue management approach to due date quotingand scheduling in an Assemble-To-Order production system. OR Spectrum 37(4) :951–982.

Gurler U, Oztop S, Sen A, 2009 Optimal bundle formation and pricing of two products with limitedstock. International Journal of Production Economics 118(2) :442 – 462.

Hahn G, Kuhn H, 2012 Simultaneous investment, operations, and financial planning in supplychains : A value-based optimization approach. International Journal of Production Economics140(2) :559 – 569.

131

Harris FHD, Pinder JP, 1995 A revenue management approach to demand management and orderbooking in Assemble-To-Order manufacturing. Journal of Operations Management 13(4) :299–309.

Hofmann E, Beck P, Füger E, 2013 CM1 : Customer segmentation and supply chain strategy.The Supply Chain Differentiation Guide, 47–79 (Springer Berlin Heidelberg), ISBN 978-3-642-31935-8.

Hulthén H, Naslund D, Norrman A, 2017 Challenges of measuring performance of the sales andoperations planning process. Operations and Supply Chain Management 10 :4–16.

Hung YF, Tsai PH, Wu GH, 2014 Application extensions from the stochastic capacity rationingdecision approach. International Journal of Production Research 1–16.

Hussain M, Khan M, Sabir H, 2016 Analysis of capacity constraints on the backlog bullwhip effectin the two-tier supply chain : a Taguchi approach. International Journal of Logistics Research andApplications 19(1) :41 – 61.

IBFP, APICS, 2011 Sales and operations planning practices and challenges. URLhttp://www.apics.org/docs/default-source/default-document-library/

apics-2011-sales-and-operations-planning-practices-and-challenges.pdf?

sfvrsn=0, accessed : 2018-05-25.

Kaipia R, Korhonen H, Hartiala H, 2006 Planning nervousness in a demand supply network : anempirical study. The International Journal of Logistics Management 17(1) :95–113.

Kibira D, Shao G, Johansson B, 2016 Framework for standardization of simulation integrated pro-duction planning. Proceedings - Winter Simulation Conference.

Kim S, Bell PC, 2015 A note on the optimal pricing and production decisions with price-drivensubstitution. International Transactions in Operational Research 22(6) :1097–1116.

Kim SW, Bell PC, 2011 Optimal pricing and production decisions in the presence of symmetricaland asymmetrical substitution. Omega 39(5) :528 – 538.

Kleijnen JP, 2017 Regression and kriging metamodels with their experimental designs in simulation :A review. European Journal of Operational Research 256(1) :1 – 16.

Kleijnen JPC, 2015 Design and analysis of simulation experiments (Springer), 2nd edition edition.

Kleijnen JPC, Sanchez SM, Lucas TW, Cioppa TM, 2005 State-of-the-art review : A user's guideto the brave new world of designing simulation experiments. INFORMS Journal on Computing17(3) :263–289.

132

Knoblich K, Heavey C, Williams P, 2015 Quantitative analysis of semiconductor supply chaincontracts with order flexibility under demand uncertainty : A case study. Computers and IndustrialEngineering 87 :394 – 406.

Kolisch R, Zatta D, 2011 Implementation of revenue management in the process industry of northamerica and europe. Journal of Revenue and Pricing Management 11(2) :191.

Kristensen J, Jonsson P, 2018 Context-based sales and operations planning research : A literature re-view and future agenda. International Journal of Physical Distribution and Logistics Management.

Lang JC, 2010 Production and inventory management with substitutions. Lecture notes in econo-mics and mathematical systems (Berlin : Springer), ISBN 3642042473 AND 9783642042478AND 978-3-642-04246-1.

Law AM, 2015 Simulation modeling and analysis. McGraw-Hill series in industrial engineeringand management science (McGraw-Hill), 5th edition edition, ISBN 9780073401324.

Law AM, Kelton WD, 2000 Simulation modeling and analysis. McGraw-Hill series in industrialengineering and management science (Boston : McGraw-Hill), 3rd edition edition.

Lehoux N, Amours SD, Frein Y, Langevin A, Penz B, 2010 Collaboration for a two-echelon sup-ply chain in the pulp and paper industry : the use of incentives to increase profit. Journal of theOperational Research Society 62(4) :581.

Lemieux S, D' Amours S, Frayret J, Gaudreault J, 2008 Integration d'outils APS dans une simu-lation multi-agent : une application à l'industrie du bois d'oeuvre. 7th International Conference onModeling, Optimization and SIMulation, MOSIM2008, 2008.

Li L, Chen ZX, 2010 A revenue management model in BTO manufacturing over an infinite horizon.International Conference on Industrial Engineering and Engineering Management, IEEM2010,2010, 1349–1352.

Lim LL, Alpan G, Penz B, 2014 Reconciling sales and operations management with distant sup-pliers in the automotive industry : A simulation approach. International Journal of ProductionEconomics 151 :20 – 36.

Lim LL, Alpan G, Penz B, 2017 A simulation-optimization approach for sales and operations plan-ning in build-to-order industries with distant sourcing : Focus on the automotive industry. Com-puters and Industrial Engineering 112 :469 – 482.

Marier P, Bolduc S, Ben Ali M, Gaudreault J, 2014 Sales and operations planning network modelfor commodity lumber products. 10th International Conference on Modeling, Optimization andSIMulation, MOSIM2014, 2014.

133

McGaughey R, Gunasekaran A, 2007 Enterprise resource planning (erp) : Past, present and future.International Journal of Enterprise Information Systems (IJEIS) 3(3) :23–35.

Meixell MJ, 2005 The impact of setup costs, commonality, and capacity on schedule stability : Anexploratory study. International Journal of Production Economics 95(1) :95 – 107.

Mentzer JT, Myers MB, Stank TP, 2007 Handbook of Global Supply Chain Management.SAGEPublications, Inc (Thousand Oaks, CA : SAGE Publications, Inc.).

Meyr H, 2009 Customer segmentation, allocation planning and order promising in Make-To-Stockproduction. OR Spectrum 31(1) :229–256.

Montgomery DC, 2009 Design and analysis of experiments (Hoboken, NJ : Wiley), 7th editionedition.

Murray PW, Agard B, Barajas MA, 2018 Forecast of individual customer’s demand from a largeand noisy dataset. Computers and Industrial Engineering 118 :33 – 43.

Nedaei H, Mahlooji H, 2014 Joint multi-objective master production scheduling and rolling horizonpolicy analysis in Make-To-Order supply chains. International Journal of Production Research52(9) :2767 – 87.

Noroozi S, Wikner J, 2017 Sales and operations planning in the process industry : A literaturereview. International Journal of Production Economics 188(Supplement C) :139 – 155.

Olaitan OA, Geraghty J, 2013 Evaluation of production control strategies for negligible–setup,multi–product, serial lines with consideration for robustness. Journal of Manufacturing Techno-logy Management 24(3) :331–357.

Olhager J, Johansson P, 2012 Linking long-term capacity management for manufacturing and ser-vice operations. Journal of Engineering and Technology Management 29(1) :22 – 33.

Olhager J, Rudberg M, Wikner J, 2001 Long-term capacity management : Linking the perspectivesfrom manufacturing strategy and sales and operations planning. International Journal of Produc-tion Economics 69 :215–225.

Olhager J, Selldin E, 2007 Manufacturing planning and control approaches : market alignment andperformance. International Journal of Production Research 45 :1469–1484.

Oliva R, Watson N, 2011 Cross-functional alignment in supply chain planning : A case study ofsales and operations planning. Journal of Operations Management 29 :434–448.

Ovchinnikov A, Boulu-Reshef B, Pfeifer PE, 2014 Balancing acquisition and retention spendingfor firms with limited capacity. Management Science 60(8) :2002–2019.

134

Papanagnou CI, Matthews-Amune O, 2017 Coping with demand volatility in retail pharmacieswith the aid of big data exploration. Computers and Operations Research 343–354.

Pedroso CB, da Silva AL, Tate WL, 2016 Sales and operations planning : Insights from a multi-casestudy of brazilian organizations. International Journal of Production Economics 182 :213 – 229.

Petrick A, Gönsch J, Steinhardt C, Klein R, 2010 Dynamic control mechanisms for revenue mana-gement with flexible products. Computers and Operations Research 37(11) :2027–2039.

Petrick A, Steinhardt C, Gönsch J, Klein R, 2012 Using flexible products to cope with demanduncertainty in revenue management. OR Spectrum 34(1) :215–242.

Phillips RL, 2005 Pricing and revenue optimization (Stanford, Calif. : Stanford Business Books).

Pibernik R, 2005 Advanced Available-To-Promise : Classification, selected methods and require-ments for operations and inventory management. International Journal of Production Economics93–94 :239 – 252.

Pibernik R, 2006 Managing stock-outs effectively with order fulfillment systems. Journal of Manu-facturing Technology Management 17 :721–736.

Pibernik R, Yadav P, 2009 Inventory reservation and real-time order promising in a Make-To-Stocksystem. OR Spectrum 31(1) :281–307.

Ponte B, Costas J, Puche J, de la Fuente D, Pino R, 2016 Holism versus reductionism in supplychain management : An economic analysis. Decision Support Systems 86 :83 – 94.

Quante R, Fleischmann M, Meyr H, 2009 A stochastic dynamic programming approach to revenuemanagement in a Make-to-Stock production system. SSRN eLibrary 1–32.

Quante R, Meyr H, Fleischmann M, 2009 Revenue management and demand fulfillment : mat-ching applications, models, and software. OR Spectrum 31 :31–62.

Quddus MA, Hossain NUI, Mohammad M, Jaradat RM, Roni MS, 2017 Sustainable networkdesign for multi-purpose pellet processing depots under biomass supply uncertainty. Computersand Industrial Engineering 110 :462 – 483.

Rafiei R, Nourelfath M, Gaudreault J, Santa-Eulalia LA, Bouchard M, 2014 A periodic re-planning approach for demand-driven wood remanufacturing industry : a real-scale application.International Journal of Production Research 52(14) :4198–4215.

Rahdar M, Wang L, Hu G, 2018 A tri-level optimization model for inventory control with uncertaindemand and lead time. International Journal of Production Economics 195 :96 – 105.

Raza SA, 2015 An integrated approach to price differentiation and inventory decisions with demandleakage. International Journal of Production Economics 164 :105 – 117.

135

Samii AB, Pibernik R, Yadav P, 2011 An inventory reservation problem with nesting and fill rate-based performance measures. International Journal of Production Economics 133(1) :393–402.

Sandhu M, Helo P, Kristianto Y, 2013 Steel supply chain management by simulation modelling.Benchmarking : An International Journal 20(1) :45 – 61.

Santa-Eulalia L, Ait-Kadi D, D' Amours S, Frayret JM, Lemieux S, 2011 Agent-based experimen-tal investigations of the robustness of tactical planning and control policies in a softwood lumbersupply chain. Production Planning and Control 22(8) :782 – 99.

Schneider MJ, Gupta S, 2016 Forecasting sales of new and existing products using consumer re-views : A random projections approach. International Journal of Forecasting 32(2) :243 – 256.

Shi W, Liu Z, Shang J, Cui Y, 2013 Multi-criteria robust design of a JIT-based cross-docking distri-bution center for an auto parts supply chain. European Journal of Operational Research 229(3) :695– 706.

Sibdari S, Pyke DF, 2010 A competitive dynamic pricing model when demand is interdependent overtime. European Journal of Operational Research 207(1) :330 – 338.

Soderborg N, 2009 Design and analysis of computer experiments. URL https://www.

slideshare.net/Vijay_Bijaj/design-and-analysis-of-computer-experiments,accessed : 2018-05-25.

Sodhi MS, Tang CS, 2011 Determining supply requirement in the sales and operations planningprocess under demand uncertainty : A stochastic programming formulation and a spreadsheet im-plementation. Journal of the Operational Research Society 62 :526–536.

Spengler T, Rehkopf S, Volling T, 2007 Revenue management in Make-To-Order manufacturing -an application to the iron and steel industry. OR Spectrum 29 :157–71.

Stadtler H, Kilger C, 2005 Supply chain management and advanced planning : concepts, models,software and case studies (Berlin ; New York : Springer), 3rd edition.

Steinhardt C, Gönsch J, 2012 Integrated revenue management approaches for capacity control withplanned upgrades. European Journal of Operational Research 223(2) :380–391.

Taguchi G, 1987 System of experimental design : engineering methods to optimize quality and mini-mize costs (UNIPUB/Kraus International Publications).

Talluri K, Van Ryzin G, 2004 The theory and practice of revenue management (Springer).

Teguia C, Chambost V, Sanaei S, D' Amours S, Stuart P, 2017 Strategic transformation of theforest industry value chain. D' Amours S, Ouhimmou M, Audy JF, Feng F, eds., Forest valuechain optimization and sustainability, 34 (Boca Raton, FL : CRC Press), ISBN 9781498704861.

136

Thomé AMT, Scavarda LF, Fernandez NS, Scavarda AJ, 2012 Sales and operations planning : Aresearch synthesis. International Journal of Production Economics 138(1) :1 – 13.

Torbat S, Khashei M, Bijari M, 2018 A hybrid probabilistic fuzzy arima model for consumptionforecasting in commodity markets. Economic Analysis and Policy 58 :22 – 31.

Torkaman S, Ghomi SF, Karimi B, 2017 Multi-stage multi-product multi-period production plan-ning with sequence-dependent setups in closed-loop supply chain. Computers and Industrial Engi-neering 113 :602 – 613.

Tsai KM, Wang SC, 2009 Multi-site Available-To-Promise modeling for Assemble-To-Order ma-nufacturing : An illustration on TFT-LCD manufacturing. International Journal of ProductionEconomics 117 :174–184.

Tuomikangas N, Kaipia R, 2014 A coordination framework for sales and operations planning :Synthesis from the literature. International Journal of Production Economics 154 :243 – 262.

Vaeztehrani A, Modarres M, Aref S, 2015 Developing an integrated revenue management andcustomer relationship management approach in the hotel industry. Journal of Revenue and PricingManagement 14(2).

Vila D, Martel A, Beauregard R, 2006 Designing logistics networks in divergent process industries :A methodology and its application to the lumber industry. International Journal of ProductionEconomics 102(2) :358 – 378.

Villegas MA, Pedregal DJ, Trapero JR, 2018 A support vector machine for model selection in de-mand forecasting applications. Computers and Industrial Engineering 1–7.

Volling T, Eren Akyol D, Wittek K, Spengler TS, 2011 A two-stage bid-price control for Make-To-Order revenue management. Computers and Operations Research 39(5) :1021–1032.

Volling T, Spengler TS, 2011 Modeling and simulation of order-driven planning policies in Build-To-Order automobile production. International Journal of Production Economics 131(1) :183 –193, innsbruck 2008.

Wagner SM, Ullrich KK, Transchel S, 2014 The game plan for aligning the organization. BusinessHorizons 57(2) :189 – 201.

Wery J, Gaudreault J, Thomas A, Marier P, 2018 Simulation-optimisation based framework forsales and operations planning taking into account new products opportunities in a co-productioncontext. Computers in Industry 94 :41 – 51.

Wochner S, Grunow M, Staeblein T, Stolletz R, 2016 Planning for ramp-ups and new productintroductions in the automotive industry : Extending sales and operations planning. InternationalJournal of Production Economics 182 :372 – 383.

137

Wu Y, J Gao J, R Yu L, 2009 Demand chain management - the new source of profit increase. IndustrialConference on Engineering and Engineering Management, 1483–1487.

Xiaodong S, Peng T, Yun C, Yue J, 2007 Revenue management for two quality level products.3rd International Conference on Wireless Communications, Networking, and Mobile Computing,WiCOM2007, 2007, 5079–82.

Yao Y, Kohli R, Sherer SA, Cederlund J, 2013 Learning curves in collaborative planning, forecas-ting, and replenishment information systems : An empirical analysis from a mobile phone manu-facturer. Journal of Operations Management 31(6) :285 – 297.

Z Wang J, Hsu PY, 2010 Advanced sales and operations planning based on integration of physicaland financial flows. International Conference on Industrial Engineering and Engineering Mana-gement, 70–74.

Zatta D, 2016 Revenue Management in Manufacturing State of the Art, Application and Profit Im-pact in the Process Industry (Springer International Publishing), ISBN 9783319302409 AND9783319302393.

Zhang Z, Wu LW, Jin C, 2006 The simulation research on inventory control of distribution systembased on customer segmentation. International Conference on Management Science and Enginee-ring, 2006, volume vol. 1, 203–7 (IEEE).

138