The Impact of SKU and Network Complexity on Inventory Levels by Joseph McCord B.S. Supply Chain Management & B.S. International Business Robert H. Smith School of Business, University of Maryland, College Park, 2009 David Novoa Garnica Dipl6me D'Ingenieur Universit6 Pierre et Marie Curie, Paris VI / Polytech Paris-UPMC, 2009 SUBMITTED TO THE ENGINEERING SYSTEMS DIVISION IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF ENGINEERING IN LOGISTICS at the ARCHNiES MASSACHUSETTS INSTITUTE OF I T CHNOLOWY MASSACHUSETTS INSTITUTE OF TECHNOLOGY JUL 162015 June 2015 LIBRARIES C 2015 Joseph McCord and David Novoa Garnica. All rights reserved. The authors hereby grant to MIT permission to reproduce and to distribute publicly paper and electronic copies of this thesis document in whole or in part in any medium now known or hereafter crated. Signature of Author.... Signature redacted Master of Engineering in Logistics Program, Engineering Systems Division .8. 20.15 ay ......................... Signature of Author . in Logistics Program, Engineering Systems Division Master of Engineig redacted Signature Signature - rd Z iMay a th, c 2015te d ........................................... Cetfe dbby.... .... rii C 8 Dr. Bruce C. Arntzen Executive Dir tor, Supply Chain Management Program ,Z 7Thesis Supervisor Acepedy......Signature Accepted by ............. S i n t r redacted e a t d 7 1 ......................................... iDr. Yossi Sheffi Director, Center for Transportation and Logistics Elisha Gray II Professor of Engineering Systems Professor, Civil and Environmental Engineering The Impact of SKU and Network Complexity on Inventory Levels by Joseph Cole McCord and David Novoa Garnica Submitted to the Engineering Systems Division On May 8 th, 2015 in Partial Fulfillment of the Requirements for the Degree of Master of Engineering in Logistics Abstract Many firms introduce new distinct products more quickly than they remove old products, and some firms have also established larger distribution networks to increase service levels or support new markets. This research applies ordinary least-squares regression and a simulation approach to identify the relationship between increased complexity and inventory levels relative to demand for a major fast-moving consumer goods company. For this research complexity is defined as the number of SKUs in a brand and the number of stocking locations for an SKU. We find that while increased complexity does translate into increased demand variability, it does not correspond to higher inventory levels. While this research does not isolate the exact reason for this disconnect, it could relate to the degree to which inventory targets recommended by an optimization software are adhered to by planning staff. For similar companies which are navigating inventory cost and complexity pressures, the research implies that there may not be a direct relationship if the company does not strictly execute an inventory policy which bases safety stock levels on forecast error. Thesis Supervisor: Dr. Bruce Arntzen Title: Executive Director, Supply Chain Management Program 2 Acknowledgements The authors would like to acknowledge Dr. Bruce Arntzen for his guidance and oversight of this research effort, staff at Unilever for their insight and feedback, and their SCM cohorts for an awesome year. Joe & David I'd like to thank Mom and Dad, Pop-Pop, Grandma, Liz, Jackie, and Pablo for their love and support in getting me to Cambridge and throughout the year. Joseph McCord I would like to take this opportunity to thank my family and friends for their encouragement and support: Anais, Melbis, Soledad, Mama, Papa, Camilo, Juan Pablo, Farid, Serge, Nathalie, Yohann, Uncles, Aunts and all the others. David Novoa Garnica 3 Table of Contents Abstract ......................................................................................................................................................... 2 Acknow ledge m ents ....................................................................................................................................... 3 Ta ble of Contents .......................................................................................................................................... 4 Tables and Figures.........................................................................................................................................6 1. Introduction...............................................................................................................................................7 2. Literature Review ...................................................................................................................................... 9 2.1 Supply chain com plexity......................................................................................................................9 2.2 SKU and network com plexity ....................................................................................................... 11 3 M ethodology ............................................................................................................................................ 13 3.1 General Experim ental Design ............................................................................................................ 13 3.1.1 Unilever Operational Context ..................................................................................................... 13 3.1.2 General Approach .......................................................................................................................... 14 3.2 Data Sources and Initial M anipulation......................................................................................... 14 3.2.1 Data Sources...................................................................................................................................14 3.2.2 Initial Data M anipulations .............................................................................................................. 16 3.3 Applied Definitions ............................................................................................................................ 17 3.4 SKU Com plexity ................................................................................................................................. 18 3.5 Network Com plexity..........................................................................................................................19 3.6 SKU and Netw ork Com plexity Sim ulation ..................................................................................... 19 4.0 Data Analysis and Results......................................................................................................................21 4.1 Dataset Sum m ary Statistics...............................................................................................................21 4.2 SKU Com plexity ................................................................................................................................. 25 4.3 Network Com plexity..........................................................................................................................27 4.4 Sim ulation Exercise ........................................................................................................................... 30 4.4.1 Simulation Scenario: Inventory Managed Weekly by Safety Stock Equation ............................ 30 4.4.2 Simulation Scenario: Inventory Managed over Lead Time Using Safety Stock Equation .......... 32 4.4.3 Sim ulation Scenario: Inventory M anaged Using "ABC" M ethod ............................................... 5. Discussion ................................................................................................................................................ 5.1 Original Research Question and Context ..................................................................................... 33 36 36 5.2 Sum m ary of Outcom es......................................................................................................................37 5.3 Im plications of Findings.....................................................................................................................38 5.4 Lim itations ......................................................................................................................................... 4 38 6 . C o nclu sio n ............................................................................................................................................... 40 R efe re n ce s ................................................................................................................................................... 42 5 Tables and Figures 16 Table 3-1. Cluster 2 M arket Dataset Sum m ary ....................................................................................... 17 Table 3-2. A pplied D efinitions ..................................................................................................................... 21 Table 4-1. Sum m ary of SKU-related Statistics....................................................................................... 22 Figure 4-1. Distribution of Demand by SKU, Cluster 1 dataset .............................................................. Figure 4-2. Distribution of Demand by SKU, Cluster 2 Category A..........................................................23 23 Table 4-2. Sum m ary of Network-related Statistics ................................................................................ Figure 4-4. Number of SKUs per Number of Nodes in Network, Cluster 1 Dataset...............................24 Figure 4-5. Number of SKUs per Number of Nodes in Network, Cluster 2 Category A Dataset.............25 Figure 4-6. Relationship between SKU Complexity and Stock Levels, Cluster 1 October 2014..............26 Figure 4-7. Relationship between SKU Complexity and Stock Levels, Cluster 2 Category B March 2015 ..26 Figure 4-8. Relationship between SKU Complexity and Stock Levels, Cluster 2 Category A April 2015.....27 Figure 4-9. Relationship between SKU Complexity and Stock Levels, Cluster 1 Category C April 2015.....27 Figure 4-10. Relationship between Network Complexity and Stock Levels, Cluster 1 October 2014 ........ 28 Figure 4-11. Relationship between Network Complexity and Stock Levels, Cluster 2 Category B March 29 2 0 1 5 ............................................................................................................................................................. Figure 4-12. Relationship between Network Complexity and Stock Levels, Cluster 2 Category A April 2015 29 ..................................................................................................................................................................... Figure 4-13. Relationship between Network Complexity and Stock Levels, Cluster 1 Category C March 30 2 0 1 5 ............................................................................................................................................................. Figure 4-14. Simple Base Stock Policy Simulation Weighted Days of Stock per Brand ........................... 31 32 Figure 4-15. Simple Base Stock Policy Simulation Nodes per SKU .......................................................... 32 Stock ............................. of Days Weighted Simulation Time Lead Figure 4-16. Base Stock Policy with Figure 4-17. Base Stock Policy with Lead Time Simulation Number of Nodes........................................33 34 Figure 4-18. 'ABC' Policy Weighted Days of Stock .................................................................................. 35 Figure 4-19. 'ABC' Policy Days of Stock against Number of Nodes per SKU .......................................... 38 Figure 5-1. Comparison of Actual and Simulated Inventory Levels ....................................................... 6 1. Introduction Any firm involved in the buying and selling of physical goods holds inventory for several reasons. Inventory allows firms to meet the general demand of consumers and hedge against uncertainty in demand or delivery lead times. While firms prefer to hold the minimum amount of inventory possible while still meeting service goals, increases in supply chain complexity have the potential to raise inventory requirements for equivalent service levels. Firms might increase the number of stockkeeping units (SKUs) or the number of network locations to capture new markets, but these decisions increase supply chain complexity and potentially inventory requirements to cover the added uncertainty of these new markets. If firms better understood the approximate inventory increase attributable to these two types of changes they would be better equipped to weigh the costs and benefits of adding new SKUs or adding network locations. Instead of incurring higher inventory costs as complexity rises, firms would have the ability to identify the point at which adding new SKUs or network locations is not worth the associated inventory cost increases. Given the ubiquity of inventory management pressures, this research is of broad interest across firms, sectors and industries, particularly in low-margin, high-volume sectors such as consumer packaged goods. Any firm which seeks to carefully manage inventory or supply chain complexity would find commercial value in an understanding of the real world relationship between the two. This research applied ordinary least-squares regression to cleaned historical datasets to identify relationships between average point-in-time inventory levels and the number of distinct stock keeping units in a brand or the number of stocking locations in the distribution network. To complement the initial findings of this approach and deepen the understanding of the implications, the research team also conducted a simulation based on hypothetical data. This simulation incorporated hypothetical measurements of demand variability and inventory levels calculated under several inventory control scenarios. Initial results of the ordinary least squares regression indicated no correlation between either form of complexity and inventory levels. However, subsequent iterations of this approach using data from 7 different operating markets produced modest indications of correlation. Examination of demand variability figures in these datasets together with the simulation exercise showed that while higher complexity does directly correlate to higher demand variability, this increased demand variability does not directly translate to higher inventory levels. These results have several implications. Firstly, as indicated by inventory theory, increased complexity does correspond to higher demand variability. However, the failure of this increased variability to directly relate to higher inventory levels could be caused by use of inventory management heuristics by planning staff. Instead of adhering to inventory targets calculated by an inventory optimization module, staff may prefer to respond to the complexity of managing hundreds of distinct products by applying non-safety stock-based approaches such as the "ABC" method. The use of these approaches instead of optimized values may effectively be the cost of increased complexity, although this research does not indicate the exact point at which complexity might cause planning staff to prefer simpler inventory control methods. This would also imply that when organizations apply inventory control methods that do not incorporate variability factors they are effectively insulated against the inventory effects of increased complexity on variability, although of course they are not insulated against decreases in inventory fill rate when variability increases. 8 2. Literature Review Through research and analysis, we sought to answer the following question: what are the effects of network and SKU complexity on inventory levels? Answering this question requires first identifying quantitative definitions of complexity as they apply to networks and categories of SKUs in a consumer packaged goods (CPG) environment. Hypothesis development also requires a theoretical understanding of how these factors should interact according to existing literature. These elements contribute to the experimental design behind the research. A literature review contributed to this research in several ways. First, it provided theoretical backgrounds to definitions used, ensuring that the research aligns with the general literature on the topic. It also provides guidance on potential experimental design approaches for inventory-related research. Finally, the literature review contributed to our understanding of the current theoretical models related to the relationship between supply chain complexity and inventory levels, which allowed the researchers to formulate better hypotheses and inform the analysis approach. Based on the research question itself, this review focuses on the following areas: the impacts of supply chain complexity generally, SKU complexity (or proliferation) and its effects, and network complexity and its effects. Additionally, the review includes formal definitions of network and SKU complexity as well as less scholarly articles from trade magazines which look at the topic of supply chain complexity. 2.1 Supply chain complexity Supply chains naturally incorporate the possibility for complexity. Although industries differ widely, the global spread of physical, informational, and financial flows ties together numerous firms and products in complicated ways. The exact nature of this complexity, however, can take numerous forms. Serdarasan, for example, identifies two broad categories of supply chain complexity: "static" - which covers supply chain structure and component variety, and "dynamic" which addresses uncertainty and randomness in supply chain activities (2013). De Leeuw further elaborates on these two broad categories to note that "static complexity refers to the expected amount of information...necessary to describe the 9 state of a planned system," while dynamic complexity relates to a supply chain's tendency to deviate from schedules (de Leeuw, Grotenhuis, and van Goor 2013). Seen this way, complexity in a supply chain can arise from the volume of the activity taking place (such as number of partners, products, or locations) as well as from the variety and unpredictable nature of these entities. For example, a supply chain that manages 100 products would require more planning than one which handles 50, but additionally a supply chain that manages 100 unique, diverse products would require more planning and management attention than one which handles 100 relatively homogenous products. Additionally, however, there are alternative ways to categorize and define forms of supply chain complexity. De Leeuw specifically identifies eight drivers of supply chain complexity drawn from prior research which he applies to the development of a case study exercise. These include scaled indicators of: upstream reliability, product diversity, size of orders, demand variability, management structure, speed of response required, lack of information synchronization, and lack of cooperation between partners (De Leeuw, 2013). These variables capture a range of perspectives on complexity which supported a quantitative comparison of existing companies. Jacobs also proposes an empirical definition of supply chain complexity which is designed to support assessments of complexity at the product portfolio, company, or supply chain levels and cover static and dynamic elements. He draws on similar histories of complexity research to develop a "General Complexity Index" which produces a comparable indicator from three components: "multiplicity," which considers the number of variants (SKUs or number of plants for example), "diversity" of those variants, and "interconnectedness" of the variants (2013). Despite their efforts to craft new empirical definitions, one common conclusion of Jacobs, De Leeuw, and Serdarasan is that there is no universally accepted definition within academic literature of supply chain complexity in general. Specifically, De Leeuw claims that the existing publications "show that there is no commonly accepted or shared measure of supply chain complexity, although all studies argue that supply chain complexity at least varies with uncertainty, variety and size" (2013). The lack of a 10 standard definition shows that there is space to tailor versions of a measurement for complexity to specific problems. 2.2 SKU and network complexity While supply chain complexity can cover numerous dimensions, SKU complexity (also called proliferation) is a particular challenge for fast-moving consumer goods companies. Companies which produce consumer products may feel pressure to develop new offerings to capture greater market share, even if it adds new complexity. Specifically, Berman identifies seven predominant causes of product proliferation which range from product offerings being a barrier to entry to managerial reluctance to prune old products (2011). Fortunately, firms do acknowledge SKU proliferation as being a significant problem. One consulting firm found that "40% of retailers identified the impact of SKU proliferation as one of the three top supply chain challenges they face" (Berman, 2011). One problematic effect of SKU proliferation is the increase in the number of "slow-movers" - products that have drastically lower sales than the average. These slow movers may "account for 40% of total items and less than 12% of sales" in the fast-moving consumer goods industry (Garry, 2011). Burke notes that SKU proliferation is a longstanding trend for many retailers. For example, "in the 1950s, a typical grocery store had about 5,000 items. In the 1990s, a supermarket carried about 30,000 items. Today, if you walk into a mass retail store like Wal-Mart, they may have over 100,000 different items" (Burke, 2007). The strategic importance and expected durability of this trend indicate a strong potential value for a model which predicts the inventory impacts of this SKU proliferation. Several sources investigate the effect of SKU proliferation. Wan, Evers, and Dresner conducted analysis of the effect of SKU proliferation (measured as the number of SKUs sold at various soft drink bottler distribution centers) on operational performance (measured by fill rate) and sales using historical records (2012). Their study finds that operational performance decreased with increases in the number of SKUs, while sales volume initially increased but then began to decrease after a peak. This phenomenon is also identified within marketing and finance literature as a "whale curve" of cumulative product (as well 11 as customer) profitability, in which the marginal profitability of successive new product introduction eventually erodes (Sievanen, Suomala, and Paranko 2004). Rossetti and Liu developed a simulation model to capture the effects of two types of SKU proliferation on operational performance (measured by fill rate) within a network of hospitals (2009). Specifically, their model compares the effects of introducing a product which is new to a particular hospital but already exists within the wider hospital system versus introducing a product which does not exist within the broader hospital system. Their simulation model shows that if SKU proliferation is comprised predominantly of products that are entirely new to the system, fill rates trend downwards. These studies support initial hypothesis development regarding SKU proliferation and inventory levels, and offer examples of empirical definitions of complexity in action. Network complexity has also benefitted from targeted research. Numbers of stocking locations, numbers of distribution tiers or echelons, and relative size of inventory requirements feature strongly in inventory modeling. One common model for the effect of network complexity on inventory requirements is known as the "Square Root Law," as the theoretical inventory required for a network of 'n' facilities is divided by the square root of n if inventories for a given market are consolidated into a single stocking location (Evers, 1995). Supply chain complexity presents significant threats to performance. Specifically, SKU complexity (or proliferation) and network complexity present potential for direct impact on inventory requirements. Significant research exists on these topics, offering expansive perspectives on definitions and causes of supply chain complexity. Effort has also taken place to generate empirical, quantitative measurements of complexity, which can support field analysis. Some of these efforts have also proven direct relationships between complexity and operational performance in the field. While the existing documentation adequately supports hypothesis formulation and the anticipated analysis, no existing literature was found which indicates a known relationship between SKU proliferation and inventory requirements. 12 3 Methodology 3.1 General Experimental Design This section presents the operational context of the research partner and the approaches to investigating the research question. The methods described in this section are designed to achieve the goal of identifying the relationships between SKU complexity, network complexity, and inventory levels. The operational context provides insight to the challenges faced by the research partner and summarizes the research partner's supply chain as it concerns the research effort. The context also informs the general experimental design by proscribing the kinds of information available and the applicability of certain methods. 3.1.1 Unilever Operational Context Unilever is a consumer packaged goods (CPG) firm headquartered in the UK and the Netherlands with brand presence in 190 countries as of 2013. The firm has 14 "Billion Euro" brands as well as numerous smaller brands (Unilever 2013). The brands cover several consumer categories, including personal care (such deodorants and soaps), hair care (such as shampoos and dyes), savory foods, and others. The mix of brands and product categories sold varies greatly between country markets in order to tailor brands to local tastes. The specific formulations, sizes, and flavors offered also vary from one market to another, and within a market over time. Unilever does not manage retail sales, but produces finished products (or contracts production from contract manufacturers) for sale and distribution to wholesalers and retailers. Unilever manages production facilities and distribution centers to support these operations. The product categories and brands will often include numerous product variations, including flavors/scents, packaging, and sizing, to differentiate the brands from competitor products, capture additional market share, and appeal to retailers who are looking to offer a complete product offering to their customers. For example, the manufacturer/distributor might offer the retailer a choice between 30 varieties of men's deodorant, even if the retailer only intends to purchase and display 10 of those 30. 13 These marketing pressures cause CPG firms to produce and offer more and more distinct SKUs over time, under the assumption that any increases in product management costs are offset by greater sales and profitability. 3.1.2 General Approach The overall effort of this research is to determine whether SKU and network complexity influence inventory levels. Intuition and existing literature suggest that as the number of SKUs or stocking locations increase, the opportunity to pool variation in demand decreases, leading to increased safety stock requirements for a given level of service (Evers 1995). To test the hypothesis that increased complexity leads to increased inventory requirements within the sponsor firm's distribution network we applied ordinary least squares regression between these two factors at several points in time within a single market. The regression function was assumed to be a power function (non-linear) given the expressions for the risk pooling effect behind the hypothesis. The approach was then replicated to other markets to test the replicability of the original model. 3.2 Data Sources and Initial Manipulation To conduct the analysis the team did not generate new data on Unilever's operations, but relied on historical records accessed by Unilever staff from company servers which capture the critical elements of complexity and inventory at points in time. The types and structures of data sources available therefore governed the range of feasible methods. For the purposes of this document, direct references to specific markets and product categories have been obscured as Cluster 1, 2, etc. for markets and Category A, B, etc. for categories. 3.2.1 Data Sources Unilever provided the analysis team with the following datasets for the initial analysis of Cluster 1 market operations: 14 * Output from inventory optimization system (known as MIO - Multiechelon Inventory Optimization) including approximately 23,000 records representing a historical snapshot of the client's Cluster 1 operations as of October 4, 2014. Each record represents a single SKU held at one of several dozen plants or distributions centers (DCs) within the market, meaning that the data cover all SKUs currently managed in the system with several records for each SKU. Each record identifies the SKU number, facility number, facility echelon (plant, Pt tier DC, or 2 nd tier DC), average demand quantity, product category, and product status (active, discontinued, promotional). The system is designed to use SKU data to recommend optimal inventory levels, although the actual inventory targets are controlled by staff. Therefore in this analysis the Cluster 1 MIO dataset was used specifically as a record of historical demand. * Output from inventory records including approximately 21,000 records representing a snapshot of the client's inventory quantities for each SKU as maintained at each facility as of October 4, 2014. Attributes include the SKU number, facility number, and current inventory quantity. While the inventory optimization system also includes inventory by SKU and location in days of stock, for the Cluster 1 operations it was assumed that this information was not up to date and would be more accurate coming from a separate inventory system. * Output from sales records including approximately 264,000 records which capture quantities of SKUs shipped to retailers by week for 40 weeks leading up to the main October 2014 period for the Cluster 1 market. The research partner also provided MIO outputs for several Cluster 2 categories for several points in time. The research partner indicated that these datasets did not have to be complemented by separate inventory and sales records as the quality of these fields in the Cluster 2 MIO database was sufficient. With the availability of multiple dates, inventory levels could be averaged to avoid the potential 15 randomness of observing inventory levels at a single point in time. The Cluster 2 MIO data included the following number of records: Table 3-1. Cluster 2 Market Dataset Summary Date April 6/7 2015 April 14 2015 Category A 18,388 18,406 Category B 3,794 2,515 Category 3.2.2 Initial Data Manipulations To facilitate the analysis process, the team prepared the Cluster 1 data in the following ways: " Inventory quantities were appended to the inventory optimization software outputs using the common SKU numbers and facility codes. Unilever recommended that for any record in the optimization software which did not have a corresponding record, inventory of 0 should be assumed. * Discontinued and promotional SKUs were removed from the analysis. At Unilever's recommendation, discontinued and promotional SKUs would not follow normal inventory management practices and would obscure any relationship in place for normal SKUs. This process removed 1,750 out of 23,801 records. " A pivot table was constructed out of the inventory optimization records. This pivot table lists a single record for each SKU, along with the brand, number of SKUs in the brand, total brand demand (quantity), SKU daily demand quantity, % of brand demand represented by the SKU, number of facilities (or nodes) where the SKU was held, the number of tiers used to distribute the SKU, inventory quantity, and days of stock (DOS) for the SKU (calculated as current inventory quantity/daily demand quantity). Days of 16 stock is considered a more accurate measure of inventory levels than inventory quantity when looking across SKUs because DOS normalizes quantities by sales volumes and some SKUs may have their inventory measured in different units (packs versus cases, for example). * Total sales volumes were calculated for each SKU in the 8 weeks preceding the inventory snapshot. Any SKU without any sales figures or sales figures totaling less than 5 units was removed from the analysis. This reduced the number of active SKUs from 4,244 to 2,560. For the Cluster 2 datasets the team applied a similar approach but did not have to append inventory data from separate source or remove additional SKUs based on sales datasets. Also, because the team had access to multiple inventory snapshots for the Cluster 2 categories, these snapshots were appended to one another into the same spreadsheet with an additional field indicating their date. 3.3 Applied Definitions While comparisons between inventory levels and "complexity" form the basis of this analysis, formal quantitative definitions for these terms are required to execute ordinary least squares regression. These definitions draw on the experience of prior research as well as the practicalities of the available data. Table 3-2 summarizes the definitions developed for this research. Table 3-2. Applied Definitions Definition Rationale SKU Complexity Number of SKUs in a Brand The datasets list the brands associated with each SKU, but there is no information on SKU turnover within the brand complexity is assumed to increase with the number of - Term SKUs Network Complexity Number of unique locations used for an SKU 17 The datasets provide the inventory levels at each location where inventory is stored for that product - this approach assumes complexity increases with the number of stocking locations used, but cannot be calculated for an entire brand because individual SKUs within the brand use distinct network paths Days of Stock (SKU) Quantity of Inventory/Daily Demand Quantity Inventory levels can be normalized to reflect underlying demand, to support comparisons between SKUs Days of Stock (Brand) Sum of Inventory Quantities/Sum of Demand Quantities for all SKUs in Brand This calculation provides a weighted inventory quantity for an entire brand Inventory on Hand (quantity) Total inventory for all records (locations) of an SKU Within the inventory optimization data, each record reflects a distinct stocking of an SKU Daily Demand (quantity) Total 'global demand' for all 'tier 1' records for an SKU Within the inventory optimization dataset, the field for "global demand" reflects the average daily demand for all customer-facing distribution centers supported by that facility for that SKU, the sum of "global demand" for all plant-level facilities reflects the demand for that SKU across the market 3.4 SKU Complexity Starting with the cleaned dataset of inventory and demand records by SKU stocking location, the team calculated total days of stock per SKU and subsequently total days of stock per brand. This produced a list of each of the brands in the dataset along with the number of SKUs in that brand and the total days of stock for that brand. Each brand was plotted against these factors in a scatterplot to support regression analysis between observed inventory levels and SKU complexity. The two factors were 18 regressed using a power function rather than a normal straight line, because the hypothesized relationship between the two variables operates on a power function. For the Cluster 2 datasets which already included a usable figure for Days of Stock for each SKU location, current inventory was calculated by multiplying Days of Stock and daily demand. Then for each SKU and brand a weighted Days of Stock was calculated and regressed against brand size. 3.5 Network Complexity Starting with the cleaned dataset of inventory and demand records by SKU stocking location, the team calculated total days of stock per SKU. Subsequently the team also identified the number of unique stocking locations employed per SKU (equal to the number of records) and the number of distinct echelons or tiers used. This process produced a list of each of the SKUs in the dataset along with the total network size and depth and days of stock for each SKU. Each SKU was plotted against these points in a scatterplot to support regression analysis between observed inventory levels and network complexity. 3.6 SKU and Network Complexity Simulation To help facilitate understanding of expected outcomes the team also conducted a simulation exercise in parallel to the primary analysis. The decision to conduct this simulation was made when initial analysis did not produce expected results, and the research team wanted to test proposed explanations. The team developed a simulation to deepen understanding by showing what inventory levels would be under various assumptions. Specifically, a simulation could show what expected inventory levels would be if inventory levels were governed by policies which account for demand variance versus common heuristics which ignore variance. This exercise required construction of a dummy dataset with features similar to the Cluster 1 market SKU summary, including a similar distribution of SKUs to brands and a similar distribution of daily demand across SKUs. In total, the simulation included 2,000 SKUs spread over 50 brands. While some features of the dummy dataset mimicked actual datasets, a dummy dataset was required in order to 19 include hypothetical demand variance data not included in the actual dataset. The simulation included demand variance measurements created under several assumptions: a coefficient of variation of 0.5 for all products and increased by a factor of the square root of the number of SKUs in the brand. From this hypothetical dataset several inventory scenarios could be developed under varying assumptions of inventory control procedures, and days of stock calculations were developed using the same approach applied to SKU and network complexity analyses described above. 20 4.0 Data Analysis and Results This section presents the quantitative results of the analyses conducted to determine the link between SKU and network complexity and observed inventory levels. The data analysis begins with summary statistics for the datasets used in the analysis. Following this are the results of the simulation and linear regression efforts for each dataset, including SKU and network complexity comparisons. 4.1 Dataset Summary Statistics This section summarizes the datasets employed for the complexity comparisons. As described in section 3, each iteration of the analysis used a set of several historical data sources provided by the research partner. In each case, these data sources provide a sense of the scope of operations in a particular market and inform expectations of what the impact of complexity would be on inventory levels. For example, knowing the number of brands and associated SKUs in a market gives a sense of the overall complexity in the product offering as well as the relative complexity within individual brands. Table 4-1 lists the SKU-related summary statistics for the datasets used and shows the comparative distributions of SKUs and their demand across the local category offerings analyzed. Table 4-1. Summary of SKU-related Statistics Cluster 1 Oct 2014 Cluster 2 Category A 2015 (Multiple Dates) No. Categories 6 1 No. Brands 77 415 No. SKUs* 2560 7368 Minimum 0 0 Maximum Mean 6,481 134 20,460 Median 24 48 Standard Deviation 310 Average Daily Demand for SKUs 191 560 *No. SKUs: number of SKUs used in the analysis after certain SKUs were removed as described in Section 3. 21 For specific markets, the distribution of demand across SKUs can be analyzed. Figure 4-1 shows the average daily demand for each SKU ranked along the X-axis by their demand quantity for the Cluster 1 dataset. 7,000 6,000 a 5,000 M E 4,000 3,000 0 a> L. Co o c0 o 00 0 C 0o 2,000 4-J o Co 0 o D0C 1,000 CD N*I It o (.0 00 0 NI '2- W 00 0D N~ Iq W 00 0 N1 It flo 00 0 SKU rank by demand Figure 4-1. Distribution of Demand by SKU, Cluster 1 dataset This distribution shows a strong skew toward a daily demand of zero units. While approximately 10% of SKUs have a daily demand of 500 units or more, more than half of the SKUs have average demand of 25 units or less per day. This distribution represents a "long tail" of a high proportion of SKUs which cumulatively represents a low proportion of total sales. 22 25000 20000 - 15000 E 10000 0 5000 4-J 0 W N n ry (Y N, c t :tqt t Lr) Lr) Lr) L) D r L r-1 t0-j 0 r-4 N r-j t.0 UL) H~ -j D .- 4 0 4 r, Lfl " W0 " C:: m TH "D m Lr) 00 H( 0 kD _4 D -4 1.0 -1 1. LO-1 M Lr 00 1H mY kw 00 -4 SKU rank by demand Figure 4-2. Distribution of Demand by SKU, Cluster 2 Category A that Again this distribution shows a strong skew towards a daily demand of zero units, showing that these demand the pattern of a "long tail" is also applicable to this category. It should be noted some SKUs do not quantities specifically represent the global demands at level one facilities, and that have a level one facility listed. Summary statistics can also be provided for the networks employed by the SKUs in a market. Table 4-2. Summary of Network-related Statistics Cluster 1 Oct 2014 Cluster 2 Category A 2015 (Multiple Dates) 53 68 1 51 58 2 11 48 3 9 34 4 6 31 5 0 14 0 0 1 4 2 I No. Unique Locations No. Locations Serving as Layer* 6 7 Minimum 23 No. Locations for an SKU * Maximum 10 12 Mean 5.1 2.5 Median 6.0 2.0 Standard Deviation 2.3 1.1 Many facilities serve as layer 1 facilities for some SKUs and as layer 2, 3, or 4 facilities for others. Table 4-2 shows the comparative network arrangements for the markets analyzed. The data show that individual SKUs can use a wide range of networks to reach customers - while some SKUs only have inventory at one facility, others may have inventory at up to 10. Figure 4-4 shows the distribution of SKUs by the number of nodes in their network for the Cluster 1 market. Again, this chart shows a significant range in the networks used by various SKUs, suggesting the possibility for complexity of operations. 800 700 600 500 4 400 6 300 Z 200 100 0 1 2 3 4 5 7 6 8 9 10 Nodes in Network Figure 4-4. Number of SKUs per Number of Nodes in Network, Cluster 24 1 Dataset 3000 2500 2000 4 6U Z 1500 1000 500 0 1 2 3 4 5 6 8 7 9 10 11 Nodes in Network Figure 4-5. Number of SKUs per Number of Nodes in Network, Cluster 2 Category A Dataset Figure 4-5 shows the distribution of network sizes for the SKUs in the Cluster 2 Category A. This distribution has a stronger skew towards lower values than the Cluster 1 categories. On average, Cluster 2 SKUs in this category flow through fewer distribution locations than Cluster 1 SKUs, even though the maximum number of location is higher for Cluster 2 Category A. 4.2 SKU Complexity This section summarizes the results of regression analysis for each dataset in the exercise, specifically for the comparisons between SKU complexity and days of stock. Figure 4-6 below shows the scatterplot, line of best fit (to a power curve) and r-squared for this comparison within the Cluster 1 dataset from October 2014. As seen in the chart, the weighted days of stock per brand do not follow the expected relationship compared to increasing complexity: the company appears to hold more inventory with a greater variance for brands with fewer SKUs, with relatively stable inventory levels for brands with more than 75 SKUs. Behind the scatterplot, brands with a small number of SKUs in some instances have a single SKU with very low demand which drives up the weighted average of the brand. For larger brands these low demand SKUs are more likely to be balanced out by SKUs with higher demand quantities. Figures 4-7 through 4-10 repeat this pattern across different time periods, markets, and categories. 25 1200 2 R = 0.047 0 1000 0 0 800 0 600 0 0 00 0* 400 OD Z- 0 150 100 50 0 200 SKUs per Brand Figure 4-6. Relationship between SKU Complexity and Stock Levels, Cluster 1 October 2014 70.00 0 60.00 R2 =0.0605 50.00 40.00 cc (A 0 0 30.00 0 20.00 PC.......... .......... 10.00 0.00 0 10 20 30 50 40 60 70 # SKUs in Brand Figure 4-7. Relationship between SKU Complexity and Stock Levels, Cluster 2 Category B March 2015 26 70.00 R2 = 0.0406 60.00 50.00 0 4 40.00 0E 30.00 m 1% 0 Z) 20.00 1W0 10.0 0.00 150 100 50 0 200 250 300 No. SKUs in Brand Figure 4-8. Relationship between SKU Complexity and Stock Levels, Cluster 2 Category A April 2015 300.00 R2 = 0.0485 0 250.00 200.00 V) 150.00 0 100.00 50.00 a w 0.00 0 50 100 150 200 250 300 # SKUs in Brand Figure 4-9. Relationship between SKU Complexity and Stock Levels, Cluster 1 Category C April 2015 4.3 Network Complexity This section summarizes the results of regression analysis for each dataset in the exercise, specifically for the comparisons between network complexity and days of stock. Similar to the comparisons of inventory levels to SKU complexity, network complexity does not appear to act as an 27 . . .... . .. .................... . . .... . . .. influencing factor. For these analyses, because individual SKUs within a brand will take different paths not make through the research partner's distribution network, measurement of network complexity does each axis. sense at a brand level. Instead, all scatterplots show individual SKUs measured against As shown in Figure 4-10, SKUs display a very wide range of inventory levels at each discrete do not follow a network size. As shown by the line of best fit, the distributions across network sizes discernable pattern. Figures 4-11 through 4-13 show this same scatterplot for varying time periods, markets, and the two factors, these categories. While they display a similar lack of the expected relationship between of nodes comparisons do show a more patterned trend of declining inventory levels as the number increases. 500 R 450 2 0.0033 8300 0 200 00 100 0 2 4 8 6 12 10 Network Size 2014 Figure 4-10. Relationship between Network Complexity and Stock Levels, Cluster 1 October 28 . .. ........ ... 160.0 R2= 0.0493 140.0 120.0 S 100.0 S 80.0 60.0 6 40.0 ". ..-.-. . 20.0 0.0 0 8 7 6 5 4 3 2 1 # Nodes in SKU Category B March 2015 Figure 4-11. Relationship between Network Complexity and Stock Levels, Cluster 2 160.00 R2 0.0732 140.00 120.00 100.00 80.00 0.0 60.00 40.00 20.00 0.00 0 2 4 6 8 10 14 12 # Nodes in SKU A April 2015 Figure 4-12. Relationship between Network Complexity and Stock Levels, Cluster 2 Category 29 . . .. .. . ...... ..... .. ... . . ........... .... .. .............. ... 700.00 R 2 = 0.0403 600.00 500.00 C 400.00 0n 0 300.00 * 4 4 200.00 0 2 4 100.00 0.00 * I 4 4 4 6 A? 8 10 I I LI: 12 14 A6 16 18 # SKUs in Brand Figure 4-13. Relationship between Network Complexity and Stock Levels, Cluster 1 Category C March 2015 4.4 Simulation Exercise Using the hypothetical dataset described in Section 3.6, scenarios were developed using varying assumptions of inventory control policies. This was conducted to show the expected outcomes if inventory levels were in fact managed according to various known inventory control policies. 4.4.1 Simulation Scenario: Inventory Managed Weekly by Safety Stock Equation The first scenario developed applied a simple base stock inventory control policy. This policy assumed a weekly review period and calculated inventory levels as cycle stock [half of weekly demand over the review period) plus safety stock (k -- a normalization factor for an assumed desired cycle service level of 0.95) times the standard deviation of demand]. Figure 4-14 shows the results of this scenario in terms of the days of stock and number of SKUs for each brand. 30 --ainval 14 R2 = - - - - .. ........................... 0.9495 12 ............... 10 2 05 8 .. -- = 3.1149xO. ' 8 0 40- 6.- .1, 0 2 200 100 0 400 300 SKUs in Brand Figure 4-14. Simple Base Stock Policy Simulation Weighted Days of Stock per Brand The number of SKUs and weighted days of stock measurements show high correlation through regression against a power function, which is to be expected because the inventory component of days of stock contains a parameter for safety stock based on the square root of the number of SKUs. This simulation presents a scenario in which inventory is highly driven by the increase in variance of demand caused by the increase in SKUs within a brand. Figure 4-15 shows a similarly clear correlation between Days of Stock and nodes per SKU. 6 R'= 0.9866 5 0.. 43 2 0 0 2 4 6 10 8 Number of Nodes 31 12 - - N-- -- -- -- -- - -I- . - - - - - - - - - - I--- - Figure 4-15. Simple Base Stock Policy Simulation Nodes per SKU 4.4.2 Simulation Scenario: Inventory Managed over Lead Time Using Safety Stock Equation The second scenario developed expanded on the first by setting the review period as the manufacturing lead time of the SKU and finding the base stock level. Figure 4-16 shows the weighted days of stock and SKUs per brand for the 50 brands in the simulation. 18 R2 = 0.5483 16 0 .. 14 .a 12 y 6 8 4 0 1 y=5.9533x 10 ....... ........ 8 0 6 4 S2 COQ 0 200 100 300 400 SKUs in Brand Figure 4-16. Base Stock Policy with Lead Time Simulation Weighted Days of Stock This simulation shows the effect of adding an additional variable to the inventory equation. The fit of the data to the power function has declined, with the r-squared value now at 0.54. The effect of the number of SKUs in a brand on inventory levels has decreased with the addition of lead times. As shown in Figure 4-17, for network complexity the r-squared drops more precipitously, because the data is not aggregated up from SKUs to brands. 32 . ......... . . . . ............ ..... - - -,-. - -- - - ---- - --l-N., NNn- -:- - -0- ................................... 18 16 16 0R' = 0.0066 * 12 . 14 10 I U 8 U 6 0 0 2 4 6 10 8 12 Number of Nodes Figure 4-17. Base Stock Policy with Lead Time Simulation Number of Nodes 4.4.3 Simulation Scenario: Inventory Managed Using "ABC" Method The third scenario included an inventory policy which ignores demand variance by setting inventory targets based on relative demand volume. Known as an "ABC method," this approach might be employed in settings where planning staff have more SKUs than they are able to manage rationally, so they manage higher volume products carefully and low volume products with less attention. Typically a small subset of the high volume products are responsible for the vast majority of total demand, while the situation can majority of products are cumulatively responsible for only a limited portion of demand. This be observed in Figure 4-1 of the 'long tail' featured in the Cluster 1 dataset. In this scenario, the products target of two responsible for 70% of total demand (375 out of the 2,000 SKUs) were assumed to have a weeks of demand, the next 25% a target of three weeks of demand, while the long tail representing 5% of total demand (957 of 2,000 SKUs) were given a target based on a normal distribution with a mean of of this eight weeks of demand and a standard deviation of three weeks. Figure 4-18 shows the results scenario. 33 .... ................ . . ...... 70 * 80 60 50 * 0 -; 40 0 0 3> 0 0 100 300 200 400 SKUs in Brand Figure 4-18. 'ABC' Policy Weighted Days of Stock The policy produces a scatterplot with no overall correlation between the two factors, as the inventory policy is based entirely on demand, and not on the variance of demand. However, brands with fewer products appear to be more likely to be influenced by 'C' products with high inventory levels, while larger brands appear to be able to average 'C' products with 'A' and 'B' products. Figure 4-19 again shows the lack of correlation introduced by this type of inventory control method. Additionally, while the correlation is similarly small, the ABC method causes a wider range in days of stock to occur than the safety stock method with multiple sources of variance. -M"w - - - - -- - - 34 200 180 160 140 D 120 CL .S 100 0 0 o R2 = 0.0034 80 III I 60 0 0 40 * I 20 80 0 0 2 4 6 10 8 12 Number of Nodes Figure 4-19. 'ABC' Policy Days of Stock against Number of Nodes per SKU 35 . ...................... 5. Discussion This section interprets the previously presented results regarding the relationship between complexity and inventory levels. The findings are briefly summarized and put into context of the original research question as well as the broader research context on applied inventory science. Additionally, this section includes commentary on the limitations of this research and the extent of its broader applicability. 5.1 Original Research Question and Context The overall effort of this research was to answer the question of whether SKU or network complexity drive inventory levels. With quantitative definitions set for these factors, the research focused on two questions: " Does a having a higher number of distinct SKUs in a brand coincide with higher amounts of inventory held for the same expected service level? * Does use of a larger distribution network (measured in number of stocking locations) coincide with higher inventory held for the same expected service level? The assumed mechanism of this relationship lies in the increase in total variance caused by splitting independent sources of unpredictability (such as demand). If a firm decides to produce and sell 20 distinct SKUs for a brand which previously included 10 distinct SKUs, the coefficient of variation and the difficulty of predicting demand for each individual SKUs increases. If the firm operates any inventory control model that includes a factor for demand uncertainty, an increase in variance produces an increase in the quantity of safety stock required to meet desired service level targets. While this overall relationship is generally assumed to hold true, the client firm has not been able to identify a strong correlation previously, and a direct link between SKU proliferation and inventory levels has not been described specifically in existing literature. A replicable methodology which identifies this relationship would allow firms in any consumerbased industry to estimate the inventory-related costs of increased complexity. 36 5.2 Summary of Outcomes As described in Section 4, correlations were found for three datasets, each representing inventory carried as part of distribution operations for the research sponsor. Each dataset included inventory and demand quantities for individual SKUs managed under brands and categories. Ordinary least squares regressions for the Cluster 1 market dataset of 77 brands found no or negative correlation between the number of SKUs and days of stock held in the network. This lack of correlation was driven by extremely high days of stock held for low demand, or "long tail" products, which for brands with fewer SKUs played a comparatively larger role. This pattern was repeated in similar analyses for separate datasets covering different time periods, markets, and categories. Least squares regression analysis of SKUs comparing inventory levels to network complexity found a similar lack of explanatory power. SKUs across time periods, markets, and categories showed significant variance in their inventory levels even at the same number of network locations. Additionally, a simulation was created to gain insight into the deviations observed between actual and hypothesized results. A simulation of a theoretical "ABC" inventory control method which ignores demand variation generated a distribution with a similar pattern to the Cluster 1 dataset. This observation, combined with knowledge of operating practices in this market indicate that planners set inventory levels in a manner closer to the "ABC" method than to optimal inventory models. Figure 5-1 provides a graphic comparison between the actual results from the Cluster 1 dataset and the simulated outcomes. From left to right, the figures include actual comparison of number of SKUs per brand and days of stock, simulated days of stock under an inventory control system where variance in demand is only driven by the number of SKUs, and simulated results under an ABC inventory control method. 37 so 014 e70r ltd n 60 11 840 105 ~40 30 10 CA) 10 0 5. mliainso inig 1 0100 0 )L0 0 0) 0 0 SKUs in Brand 20D 10D1 00 200 100 300 S~th Levels and Simulated Inventory 5-1. Comparison of Actual Figure aibliy4oeete ohgerdmn opeiydeaorepn 300 400 400 N~WSICUnnBrand i alr fti inase 5.3 Implications of Findings These results have several implications. Firstly, as indicated by inventory theory, increased complexity does correspond to higher demand variability. However, the failure of this increased variability to directly relate to higher inventory levels could be caused by use of inventory management heuristics by planning staff. Instead of adhering to inventory targets calculated by an inventory optimization module, staff may prefer to respond to the complexity of managing hundreds of distinct products by applying non-safety stock-based approaches such as the "ABC" method. The use of these approaches instead of optimized values may effectively be the cost of increased complexity, although this research does not indicate the exact point at which complexity might cause planning staff to prefer simpler inventory control methods. This would also imply that when organizations apply inventory control methods that do not incorporate variability factors they are effectively insulated against the inventory effects of increased complexity on variability. 5.4 Limitations Several factors limit the ability to extend the results of this research. Firstly, while the data analyzed represent multiple geographies and product categories, they only represent the experiences of one firm within the consumer packaged goods industry. Other firms in the same industry or firns in other industries may operate more centralized inventory control methods which would affect the relationship between complexity and inventory levels in more direct ways. The independence of demand for products within a brand would also vary between industries and would influence the variability of demand associated with addition of SKUs or network locations to a brand. 38 ............. . ... ........ - -------- Also, in complex settings, there could be many sources of complexity and variability. This research focused on changes in inventory levels attributable to SKU and network complexity, and regressed SKU levels and inventory levels against an equation which assumed independence of demand between products. Other unexplored sources of complexity which could drive inventory levels could be production or sourcing lead time variance, frequency of product mix change, and overall variance of demand within brands. - Additionally, the data used for this research reflects operations within a limited time period specifically from October 2014 to April 2015. Potentially this time period reflects unusual demand or inventory patterns for the firm which are not typical. 39 6. Conclusion This research focused on the role of complexity in driving inventory levels at a global fastmoving consumer goods firm. Specifically, it sought to answer the questions of whether increases in SKU and network complexity correspond to increases in the days of stock held. In summary, the research effort found that these specific forms of complexity do not strongly correspond to associated increases in inventory levels. Instead, changes in demand variability associated with the increased complexity do not translate into increases in actual inventory levels, and the inventory levels held follow expected patterns under inventory control models which do not include demand variance as a factor. This implies that there is no direct effect of complexity on inventory levels. While the observed inventory levels are similar to expected levels under the simulations conducted for this research, the research partner should undertake internal efforts to determine the exact methodologies used by planners to set inventory targets. This could inform change management approaches to encourage planners to use recommended optimal inventory levels more frequently. Several related questions remain uninvestigated in this research. One important factor related to the concept of the cost of complexity could be the impact on service levels - if inventory levels remain unchanged even when variability increases, the assumption would follow that associated service levels would decline. This could then translate into a useful "cost" measure when looking at the impact of complexity. It may also be useful, through difficult, to determine at what level of complexity staff prefer to apply simpler heuristics over optimal inventory management methods. Additionally, another way to examine "cost" might be to look at the difference between actual and ideal inventory levels under increasing levels of complexity. This could provide another perspective of the impact of increasing variability. Another line of valuable insight could come from investigating the role of other forms of complexity in this context. Other forms of complexity could include overall variance of demand within 40 product categories or brands, frequency of product turnover within brands, or variability between customer demand patterns. Additional research could also investigate these relationships in other industries. Potentially, in sectors with greater centralization and automation of inventory targets, more direct relationships between complexity and inventory levels might exist. 41 References Berman, Barry. (2011). "Strategies to Reduce Product Proliferation." Business Horizons 54 (6): 551-61. doi:10.1016/j.bushor.2011.07.003. Burke, R. (2007). Turning shoppers into buyers. Bloomington, IN: Kelley School of Business. & Cook, Miles. 2001. "The Complexity of Managing Complexity. (cover Story)." Transportation Distribution42 (4): 29. De Leeuw, Sander, Ruud Grotenhuis, and Ad R. van Goor. (2013). "Assessing Complexity of Supply Chains: Evidence from Wholesalers." InternationalJournalof Operations& Production Management 33 (8): 960-80. Evers, Philip T. (1995). "Expanding the Square Root Law: An Analysis of Both Safety and Cycle Stocks." Logistics and TransportationReview 31 (1): 1. Garry, Michael. (2011). "Distributors Grapple With Slow-Moving Items." SN: Supermarket News 59 (7): 32-32. Huang, Shui-Mu, and Jack C. P. Su. (2013). "Impact of Product Proliferation on the Reverse Supply Chain." Omega 41 (3): 626-39. doi:10.1016/j.omega.2012.08.003. Jacobs, Mark A. (2013). "Complexity: Toward an Empirical Measure." Technovation 33 (4-5): 111-18. doi: 10.101 6/j.technovation.2013.01.001. Rossetti, Manuel D., and Yanchao Liu. (2009). "Simulating SKU Proliferation in a Health Care Supply Chain." In Winter Simulation Conference, 2365-74. WSC '09. Austin, Texas: Winter Simulation Conference. http://dl.acm.org/citation.cfm?id=1995456.1995779. Serdarasan, Seyda. (2013). "A Review of Supply Chain Complexity Drivers." Computers & Industrial Engineering,Special Issue: The International Conferences on Computers and Industrial Engineering (ICC&IEs) - series 41, 66 (3): 533-40. doi:10.1016/j.cie.2012.12.008. Sieviinen, M., Suomala, P., & Paranko, J. (2004). Product profitability: Causes and effects. Industrial Marketing Management, 33(5), 393-401. http://doi.org/10.1016/j.indmarman.2003.08.017 Wan, Xiang, Philip T. Evers, and Martin E. Dresner. (2012). "Too Much of a Good Thing: The Impact of Product Variety on Operations and Sales Performance." Journalof Operations Management 30 (4): 316-24. doi:10.1016/j.jom.2011.12.002. 42