Moison_2019

Joint Replenishment and Base Stock Model for the U.S. Beer Industry By Nathan George Moison Bachelor of Science, Business Administration, University of South Carolina, 2015 SUBMITTED TO THE PROGRAM IN SUPPLY CHAIN MANAGEMENT IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF APPLIED SCIENCE IN SUPPLY CHAIN MANAGEMENT AT THE MASSACHUSETTS INSTITUTE OF TECHNOLOGY JUNE 2019 © 2019 Nathan George Moison. All rights reserved. The authors hereby grant to MIT permission to reproduce and to distribute publicly paper and electronic copies of this capstone document in whole or in part in any medium now known or hereafter created. Signature of Author: ____________________________________________________________________ Nathan George Moison Department of Supply Chain Management May 10, 2019 Certified by: __________________________________________________________________________ Dr. Josu é C. Vel á zquez Mart í nez Executive Director, Supply Chain Management Blended Program Capstone Advisor Certified by: __________________________________________________________________________ Dr. Sergio Caballero Research Scientist, Center for Transportation & Logistics Capstone Co-Advisor Accepted by: __________________________________________________________________________ Dr. Yossi Sheffi Director, Center for Transportation and Logistics Elisha Gray II Professor of Engineering Systems Professor, Civil and Environmental Engineering

2 Joint Replenishment and Base Stock Model for the U.S. Beer Industry by Nathan George Moison and Submitted to the Program in Supply Chain Management on May 10, 2019 in Partial Fulfillment of the Requirements for the Degree of Master of Applied Science in Supply Chain Management ABSTRACT The United States beer industry operates, by law, under a three-tier distribution network. MillerCoors has made the strategic decision to open distribution centers across the United States to handle complexity while maintaining service levels. Opening new distribution centers, and introducing a new tier to the supply chain, requires MillerCoors to define a new inventory deployment strategy to maintain their service levels to their customers. This project provides a solution to the inventory deployment problem by determining how a joint replenishment approach combined with the base stock inventory policy improves the deployment of inventory across MillerCoors supply chain. A repeatable and scalable heuristic is developed which employs joint replenishment to determine economic production frequencies and links to the base stock policy across all echelons within the supply chain. The results showed inventory reductions across 75% of all products analyzed. The results also demonstrate the importance of distribution centers to pool demand and improve alignment of inventory deployment across the supply chain. Capstone Advisor: Dr. Josue Velazquez Title: Executive Director, Supply Chain Management Blended Program Capstone Co-Advisor: Dr. Sergio Caballero Title: Research Scientist, Center for Transportation & Logistics

4 TABLE OF CONTENTS Abstract ...................................................................................................................... 2 Acknowledgements .................................................................................................... 3 Table of Contents ....................................................................................................... 4 List of Figures ............................................................................................................ 5 List of Tables ............................................................................................................. 5 1. Introduction ........................................................................................................... 6 1.1. Problem Statement ......................................................................................... 7 1.2. Relevance ....................................................................................................... 7 1.3. Motivation ...................................................................................................... 7 2. Literature Review .................................................................................................. 8 2.1. SKU Classification ......................................................................................... 8 2.2. Analytical Hierarchy Process (AHP) ............................................................. 9 2.3. Data Envelopment Analysis (DEA) ............................................................... 12 2.4. Multi-Echelon Inventory Optimization .......................................................... 14 2.5. Joint Replenishment Problem (JRP) .............................................................. 16 2.6. Base Stock Model .......................................................................................... 17 3. Methodology ......................................................................................................... 19 4. Case Study ............................................................................................................. 23 4.1. The Company ................................................................................................. 23 4.2. The Supply Chain ........................................................................................... 23 4.3. Golden Brewery Business Unit ...................................................................... 24 4.4. “3.2” Portfolio ................................................................................................ 25 4.5. Current Process .............................................................................................. 26 5. Results and Analysis ............................................................................................. 29 6. Conclusion and Managerial Implications .............................................................. 35 References ................................................................................................................... 37

5 LIST OF FIGURES Figure 1: Methodology ................................................................................................... 19 Figure 2: Current SKU Run Strategies for “3.2” Portfolio ............................................ 28 Figure 3: Distributor Change in DOI ............................................................................. 32 Figure 4: Distributor Change in Barrels ........................................................................ 33 LIST OF TABLES Table 1: Recommended Run Strategy vs. Actual Run Strategy .................................... 30

7 1.1 Problem Statement The purpose of this capstone project is to improve the way MillerCoors allocates SKUs to production frequencies and inventory policies. More specifically, we seek to answer the question – how does a joint replenishment approach improve the way we determine inventory policy for multi-echelon supply chains? By answering that question, supply chain professionals and researchers should understand how connecting the optimal production frequency for subsets of SKUs can help form base stock inventory policies in multi-echelon supply chains. Readers will then be able to make better decisions when determining the inventory policy for new products due to the general and scalable solution we present. 1.2 Relevance Many articles and papers have been written on the topics of multi-echelon supply chains, inventory policy, data envelopment analysis (DEA) and joint replenishment problems (JRP). Many papers also exist using DEA to classify SKUs and then a linear optimization to determine the appropriate inventory policies. However, there are still avenues which have not yet been explored. A gap exists in the current body of knowledge to simultaneously optimize the production frequencies within a repeating cyclical production cadence and inventory policies across a multi-echelon supply chain. We show that a heuristic combining JRP and base stock policy improves the deployment of inventory across a multi-echelon supply chain. We further support this by including actual data from industry within this capstone project to provide insights of interest to both industry professionals as well as researchers. 1.3 Motivation The motivation of this capstone project is to address an issue faced by MillerCoors and many other consumer packaged goods (CPG) companies. As consumer preferences continually

8 change, organizations are faced with an ever-increasing quantity of SKUs to manage. As the quantity of SKUs increases, managing the inventory becomes equally as challenging. SKU count reduction efforts are often challenging, due to the revenue implications of eliminating products. Solving the inventory management problem for MillerCoors helps position the organization for success in a challenging market which demands ever-increasing service level delivery to customers. 2.0 Literature Review In this section, we review the literature relevant to our research. We begin with a discussion of the evolution of inventory classification and the key approaches, models, and limitations. We then move into the body of knowledge surrounding multi-echelon supply chain inventory policies and strategies. We selected these topics because they build the foundation on which to conduct our analysis. Based on past research, we knew which techniques were applied to similar problems and were able to combine them to solve the multi-echelon inventory problem for MillerCoors. 2.1 SKU Classification Companies and organizations have long looked for ways to classify products, parts, inventory, and other items in order to improve the efficiency of decision making. A single individual cannot effectively manage and make decisions on an individual SKU basis for the thousands of SKUs many companies maintain in their portfolios. This has become even more important than ever before in recent years due in part to the proliferation of unique products available in the marketplace. The earliest insights on classifying and prioritizing items dates back to the late 1800’s (Kenton, 2019). Vilfredo Pareto, an Italian economist, made the observation that roughly 80

10 analytical processes help to minimize the impacts of subjectivity. Due to the human involvement in the hierarchy and ranking at this stage, Flores notes that no more than 7 total criteria should be analyzed because humans struggle to efficiently and consistently rank any more than that. Flores leveraged computer software to compare the criteria at each node in the hierarchy and generate the weights for each of the key criteria. Flores also notes the importance of normalizing the values for each SKU within each of the criteria between 0 and 1. Without normalizing the performance, our view of the relative performance of dollar value and lead time would be skewed. Flores applies the outlined methodology to a set of 47 SKUs for a hospital. The analysis considers average unit cost, annual dollar usage, a critical factor, and lead time as criteria for the classification. One key foundational model for multi-criteria classification is the Ramanathan Model (Ramanathan, 2006). Ramanathan uses weighted linear optimization, however they do not normalize the scores for each of the criteria before optimizing. The objective function seeks to maximize the output and assign a score for each individual decision-making unit. In order to find the optimal score for each SKU within the data set, the user changes the objective function to maximize the score of the j th item. The approach must be repeated for each item within the data set. While simple in structure and formulation, Ramanathan’s approach is cumbersome and inefficient for large sets of SKUs. The other key model we will focus on is the Ng model (Ng, 2007). Ng noted that prior approaches to inventory classification were either too subjective or too complicated to implement in practice. As a result, they developed a linear optimization model that can be solved with or without a solver program which is easily understood by all stakeholders. Ng reduces the subjectivity of scoring criteria, a drawback of AHP modeling, by having the inventory manager

11 rank the criteria in order of importance. The inventory manager only decides that one criterion is more important than another, but not by how much. While this does not completely eliminate subjectivity, it substantially reduces it. New weights are generated each time the optimization is run and normalized between 0 and 1. This allows the model the flexibility to add or remove SKUs from the analysis and then reassign weights and classifications. After the optimization runs, the resulting weights can be sorted into their ABC classifications. Ng applies their model to the same 47 SKU dataset used by Ramanathan and others and compares the results. This model is limited because it becomes more difficult to solve as more criteria and SKUs are introduced. In addition, if the priority of a particular criteria changes, the classification of a particular SKU changes roughly 40% of the time. One can expect that this would only be magnified as additional criteria are introduced and the combination of partial averages increases. In the data used in this particular example, item number 5 can be classified as an A item, a B item, or a C item depending on the priority order. This example demonstrates that while the Ng Model is easy to understand and calculate, its results are sensitive to the number of criteria and the relative importance chosen by managers. Hadi-Vencheh and Mohamadghasemi used fuzzy set theory to determine the weights of the criteria for AHP (Hadi-Vencheh & Mohamadghasemi, 2011), They argue that fuzzy theory helps to handle the subjectivity of decision makers while ranking quantitative and qualitative factors. Terms (Very High, High, Medium, Low, Very Low) are used by decision makers to rank each SKU for each of the criteria. While compared to the Flores approach, the Fuzzy AHP can be applied to any number of criteria, limitations still exist. As the number of SKUs increases, the demand on inventory managers to rank each individual SKU increases as well.

13 each classification and thus the number of SKUs in each category will fluctuate each time the analysis is performed. One aspect of DEASort that needs to be expanded is the inclusion of qualitative criteria. Liu and Huang modify the classical DEA approach by including suggested ranges of the weights for each criteria as constraints in their linear optimization (Liu & Huang, 2006). The model, like many others, includes only quantitative criteria. Hatefi, Torabi, and Bagheri extended a prior model developed by Cook, Kress, and Seiford, which incorporates quantitative and qualitative criteria (Hatefi, Torabi, & Bagheri, 2014). Hatefi, et al also include a discrimination value, e , in their model which sets a minimum for the weights of each criteria as well as the difference between different levels of the qualitative criteria. The proposed model is applied to the same 47 SKU set as the Ramanathan model, which also extends Ramanathan’s conclusions. They also minimize the subjectivity found in prior models by using a dimensional unit vector to categorize the qualitative factors instead of a Likert scale. The number of levels in the unit vector corresponds to the values of the qualitative criteria. In their example, the qualitative factors can take the value of very critical, moderately critical, or non-critical . Therefore, the unit vector would have 3 levels. Millstein, Yang, and Li propose an extension to multi-criteria inventory classification that optimizes the number of groups in which to categorize SKUs along with their respective service levels. The methodology is then able to classify a set of SKUs into the appropriate categories (Millstein, Yang, & Li, 2014). Other approaches using Machine Learning Algorithms have emerged to support multi- criteria inventory classification. Both supervised (Lolli, Ishizaka, Gamberini, Balugani, & Rimini, 2017) and unsupervised approaches (Ladhari, Babai, & Lajili, 2013) have been used.

14 Decision trees and random forests are commonly used for these types of analyses, as they perform well when a data set needs to be sorted. Further opportunities exist to include qualitative data in the machine learning analyses. Yu analyzed different machine learning techniques including support vector machines, backpropagation networks, and k-nearest neighbors and compared them to a multiple discriminant analysis (Yu, 2011). Yu analyzed the same set from the Ramanathan study and concluded that the support vector machine approach yielded the most accurate results. 2.4 Multi-Echelon Inventory Optimization Researchers have also studied multi-echelon supply chain inventory strategies since the middle of the 20 th century. Multi-echelon inventory management, unlike individualized approaches, focuses on optimizing the entire supply chain rather than just a single stage (Sbai & Berrado, 2018). Recently, a comprehensive taxonomy paper was written for the area of researching multi-echelon inventory models. The review paper includes an organized classification of the papers based on the assumptions, research goals, and methodologies (de Kok et al., 2018). From this review, we see that opportunities remain to better understand multi- echelon supply chains with capacity constraints as well as provide research with field study- based analyses, less than 5% of papers reviewed. This capstone addresses both of those gaps in addition to incorporating SKU Classification as an input to determine the inventory policy. Li and Liu formulated a comprehensive multi-objective multi-echelon inventory model which incorporates constraints that impact the end to end supply chain. The model uses thorough production, transportation, and cost constraints to optimize all objectives (Li & Liu, 2010). The formulation of this model may be difficult to implement in the field, due to its complexity which may not be easily understood by supply chain managers. Tjhin and Pandey

16 GSM nodes, which takes advantage of the benefits of both strategies(Klosterhalfen, Dittmar, & Minner, 2013). Further extensions of GSM have looked at optimizing reorder intervals as well as the addition of new constraints (Eruguz, Jemai, Sahin, & Dallery, 2014). Researchers have started adding capacity constraints, but they pertain mainly to the volume of units the system can process at a single stage (Eruguz, Sahin, Jemai, & Dallery, 2016). More recently, studies have looked at combining multi-echelon inventory policies with segmentation by allocating products to a particular supply chain segment based on their classifications (Fichtinger, Chan, & Yates, 2017). Gupta further extends GSM and combines traditional single criteria SKU classification with a multi echelon inventory strategy to determine the optimal distribution network design (Gupta, 2008). 2.5 Joint Replenishment Problem (JRP) While the approaches highlighted above have certainly contributed greatly to the understanding of inventory policies in multi-echelon supply chains, we feel there are further opportunities using a joint replenishment problem (JRP) approach. Joint replenishment refers to the economies of scale which can be achieved by combining the replenishment of items and reducing the ordering costs. The total costs can be reduced by pooling the replenishment orders and spreading the fixed order costs across multiple products instead of a single product. The DEA approach has certain limitations which are difficult to effectively mitigate, such as subjectivity and human interaction. In addition, these approaches typically also meet challenges when trying to expand to broad product portfolios. JRP, explored as early as the 1960’s (Khouja & Goyal, 2008), addresses many of these shortcomings. The main objective of a JRP model is to determine the replenishment scheme for a group of SKUs which minimizes the overall cost.

17 This uses very similar assumptions to the economic order quantity (EOQ) model. The joint replenishment model also relates to principles from the Toyota Production system around production smoothing and mixed model scheduling. Chopra and Meindl present an algorithm, which applies the JRP principles, to calculate a near optimal ordering frequency for portfolios of SKUs which minimizes the total cost (Chopra & Meindl, 2013). The approach presents a five-step process to calculate the relative order frequencies as compared to the portfolio item which must be ordered the most frequently. Another approach takes into account not only the joint fixed costs associated with production, but also the variable product specific cost to determine an optimal solution for all product offerings (Bayindir, Birbil, & Frenk, 2006). This focused on a single stage inventory model while multi-echelon applications of this approach still need to be explored. Another JRP model focused on a single item and multiple locations, instead of multiple items and multiple locations (Silva & Gao, 2013). Other solutions include constraints for minimum order quantities, which better simulate real-world operations (Porras & Dekker, 2006). Various capacity restrictions have also been explored to solve the JRP in order to generate more practical operational solutions (Moon & Cha, 2006; Amaya, Carvajal, & Castaño, 2013; Ongkunaruk, Wahab, & Chen, 2016). 2.6 Base Stock Model The base stock model has been widely discussed in literature. Graves and Willems discuss the base stock policies for multi-echelon supply chains (Graves & Willems, 2003). The base stock model ties a target inventory level to the level of service a supply chain wishes to deliver. It balances both the cycle stock and safety stock to determine the appropriate level of inventory. They model a supply network with deterministic lead-times, stationary demand, common review

19 3. Data and Methodology In this chapter, we present the methodology used to simultaneously determine the production frequencies and the inventory policies for a multi-echelon supply chain. We begin by outlining the formulation used for the heuristic. We will then introduce a case study of MillerCoors where we apply the methodology to actual data and present our outcomes. The main steps of the methodology are as follows, which will be further elaborated throughout the remainder of this chapter. First, we must collect and clean the relevant data. We then calculate the recommended run strategies based on the joint replenishment (JRP) algorithm. This allows us to categorize the SKUs and prepare them for the inventory policy portion of the heuristic. The classification and run strategy then feed into the base stock calculation as the lead time. We calculate the appropriate inventory policy for the Distribution Center as well as all distributors. All pertinent and relevant formulas are included in section 3.1. Figure 1: Methodology 3.1 Heuristic Our analysis involves a combination of the joint replenishment process heuristic and the base stock policy. In the following section, we will define all of the equations and variables used to generate our production cadence driven inventory policies. ࠵? " # = % &’ ( ) ( *(,-. ( ) (Eq.1) Collect and Clean Data Calculate Run Strategies (JRP) Calculate Base Stock Policy

20 We will start by defining all variables within Eq.1. ℎ represents the holding cost, as a percentage of unit cost. For our analysis, we will assume a 15% holding cost. ࠵? ࠵? represents the Unit Cost for each SKU, ࠵? . ࠵? ࠵? represents the Annual Demand for each SKU, ࠵? . ࠵? represents the fixed setup cost for each production run. ࠵? ࠵? represents the unique cost for each SKU, ࠵? . For our analysis, we will interpret ࠵? 7 as the cost of complexity for each SKU based on the annual volume. Cost of complexity refers to the increased costs and decreased efficiencies that develop as the number of low volume SKUs increases. The cost is used to quantify the added burden on the supply chain for each incremental SKU added to a portfolio. Finally, we are solving for ࠵? # , which is the economic number of production runs for each SKU per year. Once we have calculated ࠵? # 7 , we identify which SKU has the largest value. This will become our reference point to determine the frequency to produce every other SKU, relative to the ࠵? # SKU. ࠵? 8 7 = % &’ ( ) ( *. ( (Eq.2) Next, in Eq.2, we calculate ࠵? ࠵? 9 , which is the economic production frequency for all SKUs which are not ࠵? # . All other variables are the same as in Eq.1. The reason that the fixed cost, S, is no longer included is because the most frequently produced SKU absorbs all of this cost. The heuristic assumes that the most frequently produce SKU is produced with every run of the other SKUs. ࠵? 7 = ; < < = 888 # > (Eq.3) Eq.3 determines the production frequency, ࠵? ࠵? , for every ࠵? 9 7 relative to ࠵? # , for each SKU, ࠵?. The brackets surrounding the equation signify us to round the result up to the nearest integer.