SKU Segmentation Strategy for a Global Retail Supply Chain By Huiping Jin MS Finance, Case Western Reserve University, 2013 Master of Business Administration, Tongji University, 2012 SMASSACHUSE[TSINSIJE OF TECHNOLOGY JUL 17 2014 And LIBRARIES_ Brad Gilligan B.S. Business Administration, Colorado State University, 2010 Submitted to the Engineering Systems Division in Partial Fulfillment of the Requirements for the Degree of Master of Engineering in Logistics at the Massachusetts Institute of Technology June 2014 02014 Bradley Michael Gilligan and Huiping Jin. 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 created. Signature of Author ............................... Signature redacted Master of Engineering in Logistics Program, Engi eering Systems Division Signature redacted Signature of A uthor ................... May 13, 2014 ..... . ........................................... Master of Engineering in Log~tics.Pr.g.,.giermn Systems Division May13,2014 Signature redacted' C ertified by ................................................. Signature Accepted by ................. r ......................... Edgar E. Blanco Thesis Supervisor Principal Research Associate Executive Director IT S ALE Network Latin America redacted MIT Center for Transportation & Logistics ........................................................................ Prof. Yossi Sheffi Professor, Engineering Systems Division Professor, Civil and Environmental Engineering Department Director, Center for Transportation and Logistics Director, Engineering Systems Division W 1 SKU Segmentation Strategy for a Global Retail Supply Chain By Brad Gilligan And Huiping Jin Submitted to the Engineering Systems Division in Partial Fulfillment of the Requirements for the Degree of Master of Engineering in Logistics at the Massachusetts Institute of Technology ABSTRACT The concept of using different supply chain strategies for different products or product families is a fairly simple component of supply chain management. This practice, known as SKU segmentation, is widely used by many companies. However, most research and success stories involve a relatively stable portfolio of brands and products, and products with easily identifiable attributes such as profit margins and demand. This thesis attempts to determine if and how a SKU segmentation can be conducted when product mix is constantly changing and many key variables used in traditional segmentations are not known in advance. To explore this problem, we analyze one year of purchase order data and shipment data provided by our sponsor company. The objective is to use data from purchase orders to predict which items are candidates for different supply chain configurations (i.e. an expedited supply chain for time-sensitive products or an efficient supply chain if there is opportunity to reduce cost and still meet demand). We start by mapping the current supply chain process using historical data and interviews with employees. The key piece of the process we want to understand is how early or late products arrive at destination in relation to when those goods are expected in retail stores (a metric we refer to as "destination dwell time"). We then use visualization and statistical analysis to determine what PO information is related to the destination dwell time. After testing various multi-factor regression models to predict the length of this dwell time, we conclude that a neural network regression model predicts this time most accurately. We then assess whether or not it is feasible for the sponsor company to use this model to "speed up" or "slow down" the supply chain for different products as needed. Thesis Supervisor: Edgar Blanco Title: 2 ACKNOWLEDGEMENTS On behalf of Brad Gilligan: I thank the entire faculty and staff of MIT's Center for Transportation and Logistics for designing and managing this challenging and rewarding program. I thank my classmates for enriching this experience with their insight and support, and my thesis partner Huiping Jin and advisor Edgar Blanco for their contributions to this project. I thoroughly enjoyed working with such brilliant people and learned so much from both of you. I thank the many helpful partners at our sponsor company for making this project possible. We appreciate you taking the time to teach us about your business, arranging tours, and providing data. It was a pleasure to work with all of you. I thank my parents, Tim and Jill Gilligan for always challenging me to better myself. You taught me the value of education and hard work, and none of my achievements would be possible without you. I also thank Leslie Herring for being by my side throughout this hectic year. Your love and support made this experience more manageable and enjoyable than I could have imagined. On behalf of Huiping Jin: First of all, I would like to express my deepest appreciation to our thesis advisor, Dr. Edgar E Blanco, for his support and motivation. His guidance helped me in completing this research and his enthusiasm in supply chain management really encouraged me to fully commit myself to this area. I would also like to give my special thanks to Dr. Bruce C. Arntzen for his guidance in getting my graduate study at MIT started on the right foot and at the right direction. In addition, my thanks also go to Jennifer Ademi and Allison Sturchio who had provided endless care to my life at MIT and valuable advice on my career. You are the ones that made my life at SCM an enjoyable experience. I am also indebted to all the faculty members at MIT ESD SCM program for giving me a lot of insights and help in completing my study and research. I would like to thank my thesis partner, Bradley M. Gilligan, who has made tremendous contribution in completing this thesis. I really enjoyed working with you in completing this research. Finally, and most importantly, I would like to thank my wife Yefei Gu, my father Yong Jin, and my mother Lianfeng Tang. It is your support and love that made me this far. I greatly appreciate for everything you have done for me. 3 CONTENTS A BSTRA CT.................................................................................................................................... 2 A CKN O W LED G EMEN TS..................................................................................................... 3 CON TEN TS..................................................................................................................................... 4 TABLE O F FIGU RES.................................................................................................................... 6 TA BLE O F TA BLES ..................................................................................................................... 8 1 Introduction.................................................................................................................................. 9 1.1 Company-Specific Challenges of SKU Segmentation................................................ 1.2 Hypothesis.....................................................11 1.3 A pproach ........................................................................................................................ 11 2 Literature Review ...................................................................................................................... 12 10 2.1 Purpose of Stock K eeping U nit (SKU) segmentation ...................................................... 12 2.2 SK U segm entation dim ensions........................................................................................ 12 2.3 SK U segm entation practice for fashion retail industry .................................................... 14 3 M ethods ..................................................................................................................................... 15 3.1 D ata Collection ................................................................................................................... 16 3.2 Initial A nalysis of D ata and Supply Chain Process ........................................................ 17 3.3 Quantitative A nalysis...................................................................................................... 20 4 D ata A nalysis and Results......................................................................................................... 21 4.1 Initial D ata Statistical D escription................................................................................... 23 4 4.1.1 M easuring Tim ing Attributes.................................................................................... 24 4.1.2 D efining and M easuring Dw ell Tim e ...................................................................... 28 4.2 H ypothesis .......................................................................................................................... 29 4.3 M odel Construction ............................................................................................................ 33 4.3.1 Ordered-Probit M ethod............................................................................................. 34 4.3.2 N eural N etw ork M odel............................................................................................. 38 4.4 Tests and Results ................................................................................................................ 45 4.4.1 Ordered-Probit Model............................................................................................... 45 4.4.2 N eural N etw ork M odel............................................................................................. 46 4.5 Potential A pplication of M odel (Proof of Concept) ........................................................ 48 5 Conclusion ................................................................................................................................. 51 A ppendix I - Ordered Probit Model Output:............................................................................. 54 Appendix II - M atlab script for neural network m odel............................................................. 56 References..................................................................................................................................... 57 5 TABLE OF FIGURES Figure 1 - Initial Understanding of Supply Chain Process and Terminology ........................... 18 Figure 2 - PO Lead Time Histogram ........................................................................................ 19 Figure 3 - Lead Time Components.......................................................................................... 22 Figure 4 - PO Attributes and Descriptions............................................................................... 24 Figure 5 - PO Create Date to Header Date Statistics (in days).................................................. 25 Figure 6 - Lead Time Components with Statistics ................................................................... 27 Figure 7 - Dwell Time Statistics (in days)................................................................................. 29 Figure 8 - A scatterplot showing the days between PO Create and Cargo Receipt on Y-axis, and Destination Dwell Time on x-axis. Different colors identify different departments, and size of circle represents order quantity................................................................................................. 31 Figure 9 - Figure 8 has been modified to show a department with more variability than normal. ...................................................................................................................................................... 32 Figure 10 - Figure 8 has been modified again to show a department with stable timing.......... 32 Figure 11 - Ordered Probit Model Output Summary............................................................... 36 Figure 12 - Ordered-Probit Prediction Evaluation.................................................................... 37 Figure 13 - Model Training and Validation............................................................................... 42 Figure 14 - Error Distribution Histogram................................................................................. 42 Figure 15 - Regress Predicted Value on Actual (In-sample test) ............................................ 43 Figure 16 - Regress Predicted Value on Actual (Out-of-Sample Prediction for 100 POs)..... 44 Figure 17 - Predicted Result VS Actual Dwell Time (Out-of-Sample Prediction for 100 POs).. 44 Figure 18 - Performance Sensitivity Relative to Change of Model Size (Number of prediction= 1,000 PO s)..................................................................................................................48 6 Figure 19 - Histogram of Actual Cargo Receipt to First Keytrol................................................. 50 Figure 20 - Histogram of Actual Dwell Time for 100 POs .......................................................... 51 Figure 21 - Histogram of Improved Dwell Time for 100 POs ..................................................... 51 7 TABLE OF TABLES Table 1 - Factors influencing supply chain segmentation ........................................................ 13 Table 2 - Ordered Probit Model - Test Result for 1,000 POs ................................................. 45 Table 3 - Test Results without updating the model (model size=3,000, gap=3,000)............... 47 Table 4 - Test Results with updating the model (model size=3,000, gap=3,000)........................ 47 Table 5 - Model Parameters Used in the Application Example ............................................... 49 Table 6 - Segment Predicted Dwell Time into Predefined Categories...................................... 49 Table 7 - Example Strategies for Supply Chain Segments........................................................ 50 8 1 Introduction Before describing SKU segmentation, it is important to clarify what is meant by the term "SKU". SKU is an abbreviation for Stock Keeping Unit and refers to a specific product within a company's catalog. As consumers, we see SKUs every day in the form of barcodes. If you walk through any large retail store, you will realize there are thousands of different SKUs, and each one can have very different characteristics such as size, price, and demand. A SKU segmentation is a simple concept that suggests there is a benefit to handling certain products differently. To illustrate this concept, consider the way you may buy and care for a suit or dress for a special event, and compare it to the way you buy socks. You would likely spend a lot of time shopping around for the suit or dress and try on several different products. You would make sure to hang it safely in your closet and would have it cleaned by a professional dry cleaner. When buying socks, you would grab the product that is easiest to find and reasonably priced, and store them and clean them with little care. The advantage of these different approaches is that you get a quality suit or dress and you do not waste too much time or effort worrying about socks. This exact same principle applied to an organization can tell them which products they should focus on and which can be handled with little effort. The advantages are satisfying customers while minimizing supply chain costs. Our research explores how this concept can be utilized by a global retailer that procures and sells a wide range of products in multiple regions. To simplify this project we focus only on items that are procured in China for sale in the USA, but the intent is for this approach to be scalable and repeatable to be applied to other parts of the business. We start by looking at their existing product mix and supply chain process, then identifying which SKUs might need to be handled in a more proactive and responsive manner, and which SKUs could be handled more economically. 9 1.1 Company-Specific Challenges of SKU Segmentation A SKU segmentation can be very straightforward and useful for companies with a mostly static product base. For example, a company with a very high-cost high-margin product line with volatile demand would obviously benefit from a flexible and responsive supply chain. They can afford to spend more on transportation in order to meet demand. On the other hand, a company that sells a low-cost everyday item with stable demand can utilize a slow and economical supply chain to control costs. In the real world, these segments of SKUs are not always so easily identifiable. What if the company's product mix changes and they do not know in advance what their margins and their demand will be? What if some SKUs are inherently more difficult to handle and require additional time in the supply chain regardless of traditional attributes? These are questions difficult to answer while working with a company that has a unique business model. Currently, our sponsor company treats every single product the same in terms of procurement, transportation, and storage. This approach can result in high-value products being delivered to stores late and missing sales, while low-value items could be delivered too early and consequently take up valuable space in a distribution center. The company sells a large variety of products, with price points that range from under ten dollars to several hundred dollars and annual sales between a single unit and over one million units. This wide range of products suggests there is definitely potential to customize the supply chain process for different SKUs, but the company's business has some unique complexities that make this segmentation challenging. To begin, the company has a very dynamic product mix and an equally dynamic pool of suppliers. Their business is seasonal, and consumer preferences change quickly and frequently. It 10 is very rare for them to sell the same SKU more than once, and their product assortment changes constantly. The supplier market is highly fragmented and competitive, so they are constantly working with new suppliers depending on where they can get products at the best value. Because a SKU segmentation relies on the attributes and behavior of suppliers, products, and customers over time, this volatility makes the process much more difficult. 1.2 Hypothesis Even with the challenges presented by this business model, there is potential to identify SKUs which can be handled differently to achieve improvements in cost and/or service. Because the products and suppliers are constantly changing, we will look at higher level attributes of SKUs such as merchandise departments, order sizes, and origin locations rather than lower level attributes such as specific suppliers and styles, the most traditional segmentation attributes. 1.3 Approach We used a combination of qualitative and quantitative analysis to identify when and how this company could treat certain SKUs differently. The qualitative piece involved working with stakeholders to clearly understand and map out their supply chain process from procurement to store delivery. The quantitative piece involved statistical analysis to predict lead times and demand for products based on SKU attributes. The deliverable is a model that could tell this company which SKUs should be expedited in order to meet demand, which could be held or processed differently to reduce costs, and which fit the current process. 11 2 Literature Review SKU segmentation is not a new or innovative idea. Research on the topic is readily available, and the technique is already used by many companies to improve supply chain performance. However, because every company has a different mix of SKUs and different feasible supply chain configurations, every segmentation requires a unique approach. Our sponsor company's business model proved to be an exception to many of the conventional rules used to segment SKUs. In this literature review, we will discuss the history and purpose of SKU segmentation, the typical and widely used methods, and industry-specific considerations. We will then describe how the traditional approach to segmentation did not work for our sponsor company. 2.1 Purpose of Stock Keeping Unit (SKU) segmentation A segmentation strategy is a systematic method of separating products into different buckets for a certain purpose, which can include maximizing market share, minimizing risk, improving efficiency, etc. The approach to SKU segmentation is usually to group SKUs based on criteria that indicate which supply chain management strategies can be used to maximize a firm's value. It can also be described as a process of understanding the nature of the demand for different SKUs and devising supply chain strategies that can best satisfy that demand (Fisher, 1997). 2.2 SKU segmentation dimensions SKUs can be segmented based on many dimensions. The most traditional way is the ABC classification method, which is to segment the SKUs based on the dollar sales volume. The underlying theory of this segmentation method is the 20-80 Pareto principle, which suggests that 12 20% of the SKUs account for roughly 80% of total revenue, and those 20% should deserve the most attention from management. The benefits of using the ABC segmentation method are that it is very easy to implement and there are many established inventory control strategies based on it, such as continuous review (s, Q) inventory strategy, period review (R, S) inventor strategy, order-up-to (s, S) inventory strategy etc. The shortcoming of this method is that it utilizes only one dimension to segment the SKUs. Under ABC segmentation method, SKUs within the same bucket can still have very different characteristics. For example, two SKUs may have similar dollar value contribution, but one may have high unit value but low demand, and the other may have low unit value but high demand. Obviously, it is not appropriate to apply the same strategy to both SKUs. Due to the limitations of the ABC method, many other dimensions have been used for segmentation. Additional product characteristics include profit margin, volume, demand volatility, etc. Anthony Lovell summarized these dimensions as shown in table 1. (Anthony, 2005) Table 1 - Factors influencing supply chain segmentation Group Product Market Factor Life cycle Variety within product group Product type: functional or innovative Handling characteristics Shelf life Physical size and weight Value PVD Demand location/dispersion Demand level (throughput) Demand variability Service expectations Limitations on raw material Source Economies of scale 13 Production flexibility Lead-time Geographic and commercial environment Existing infra-structure Transport mode availability Customs/duties/trade areas Legislation In addition to introducing new dimensions, SKUs can be segmented using multiple dimensions instead of just one or two. A common approach is to first segment the SKUs based on two dimensions, then add another dimension as a new axis to further slice the SKUs. An example could be segmenting the SKUs based on volume ani volatility first, then adding profit margin as the third axis to form a 3-dimensional segmentation. In general, segmentation dimensions are industry specific or company specific. There is no universal list of attributes that can be applied for all companies and all situations. Therefore, when developing a segmentation strategy for our sponsor company, it becomes very important to understand their industry and business model. 2.3 SKU segmentation practice for fashion retail industry The retail fashion industry, from a supply chain management perspective, is usually described as a perishable-goods industry. It is assumed that all the goods purchased during the period will be sold out by the end of the season. This assumption is quite reasonable because the industry is highly seasonal and excess inventory is usually sold via mark-downs. Therefore, the supply chain model is often characterized as a single-period model. A classic example of this is the newsvendor model, which assumes that every unit of supply will be sold or salvaged by the end of a given time period with known demand distribution and fixed cost. 14 Based on this model, SKU segmentation for this industry has focused on lost sales, demand patterns, margin, lead time, and holding cost. Different supply chain strategies are then developed to manage the flow of the goods from suppliers to retail stores. The goal is to balance the tradeoff between lost sales due to stock outs and the cost of excess inventory. On a higher level, those supply chain strategies can be categorized into several groups, such as a "responsive" supply chain focused on reacting quickly to changes in demand, or an efficient supply chain that aims to use the most economical transportation and storage options. Our sponsor company's suppliers and product mix are constantly changing. This prevents us from using many traditional dimensions used for segmentation such as demand variability. Another caveat that made the segmentation even more challenging was that the sponsor company requested that price and margin be excluded from the analysis because product availability is a critical component of their strategy, regardless of profitability. In order to segment their products, we will look at patterns and trends in the import process for different categories of products. They currently treat all items in their supply chain equally, but because they know that some items sit in their distribution centers for extended periods while others are delivered late to retail stores, they believe that certain products have more urgency than others. The objective of this thesis is to identify groups of SKUs that may warrant a different supply chain process in order to reduce costs or better meet customer demand. 3 Methods A traditional Stock Keeping Unit (SKU) segmentation begins by segmenting products based on cost, profit margin, and the volume and variability of demand. Because the sponsor company did not want to consider the cost of an item, and because they rarely sell the same SKU more than 15 once, we could not utilize profit margin or historical demand. This segmentation started by identifying items that had abnormally long lead times, then working backward to find out what characteristics could be contributing to those lead times. This process involved gathering data from the sponsor company and learning about its supply chain through site visits and interviews, plotting data to identify patterns and trends, and then building a regression model and simulations to test and validate the correlation of variables. 3.1 Data Collection The first step was a face-to-face meeting with the key stakeholders of this project (VP of International Logistics and Trade Compliance and VP of Logistics Development) at their corporate headquarters. This meeting served as an introduction to the company's supply chain process and provided an opportunity to ask questions to determine which variables could be relevant to the project. The next step was to tour facilities that receive and ship goods, in order to see the supply chain in action and observe the physical flow of materials. These facilities included a third party deconsolidator (crossdock) and a company distribution center. After developing an understanding of the company's end-to-end supply chain, we then looked at one year of historical data. In order to narrow the focus of the project, we looked at only data for shipments from China to the US. The data included 298,754 records (rows) and 48 fields (columns). Each record represented one shipment of one SKU. Shipments that contained multiple SKUs were listed in more than one record, and multiple shipments of a single SKU were also listed separately. The data included 491 unique suppliers, 38,306 unique Purchase Orders, and 33,566 unique SKUs. 16 3.2 Initial Analysis of Data and Supply Chain Process Once we received the data, we tried to tie each element of data to the specific supply chain processes that were explained by the company and witnessed at their facilities. We also clarified when each element of data was created and how it was generated (i.e. the PO Quantity is generated at the time the PO is created by the buyer). After validating the source and meaning of the data, we then attempted to validate the actual dates and numbers. After identifying records with missing and incorrect data, we made revisions before continuing the analysis. An overview of the supply chain process and definitions of key dates and terms is shown in Figure 1 on the following page. 17 ORIGIN LEAD TIME Supplier PO Create Date = Sponsor company agrees to buy goods. Start Ship Date = Sponsor company agrees to accept shipments of this PO. Book Date = Supplier arranges shipment of goods. Con Can Date = The last day the sponsor company will accept shipments of this PO. Consolidator (accepts shipments on behalf of sponsor company and arranges ocean shipping) Cargo Receipt Date = Goods are received by the consolidator. Consolidation Date = Goods are in an ocean container and ready to be shipped. Origin Port ETD = Estimated time of departure. This is the date the goods leave the port. OCEAN TRANSIT TIME DESTINATION LEAD TIME Destination Port ETA = Date the vessel reaches US port. Deconsolidator (receives containers and ships goods to sponsor company distribution centers) Arrive Decon = Goods received by deconsolidator. Distribution Center (DC) 1st Keytrol Date = Goods received by distribution center. IRetail Store LP Date = Goods expected in store. Figure 1 - Initial Understanding of Supply Chain Process and Terminology The next stage of the qualitative analysis entailed looking for high-level patterns and trends in SKU lead times. We used a range of visualizations to reveal patterns and found histograms, boxplots, and scatter-plots to be the most useful. The analysis started by looking at the total lead time from the creation of a purchase order to distribution to retail locations. The following histogram shows the distribution of this lead time. 18 130K 120K 110K 100K 90K 80K 70K 60K SOK 40K 30K 20K 10K OK Null 0 50 100 150 200 250 300 350 400 550 Days from PO Creation to Retail Store Distribution Figure 2 - PO Lead Time Histogram After looking at this total lead time, we focused on anomalies and looked at additional details to determine which specific processes were contributing to the long lead times. For example, were the long lead times a result of the time spent getting from the supplier to the origin port, or the result of time spent at the company's DC in the destination country? After identifying the outliers that could be causing unpredictability and unnecessary costs in the supply chain, we looked for factors that these records had in common in an effort to determine 19 causality. We developed theories about product departments, seasons, and other variables that could be correlated with lead time. 3.3 Quantitative Analysis Using the variables identified during the qualitative analysis, we built regression models to explore how well lead time could be predicted using specific variables. For instance, do products in certain departments or products sold in certain seasons take longer to get from origin to destination? To build the regression model, we focused on the top seventeen SKU categories by order quantity, which account for 80% of total volume. We set lead time to be the dependent variable and ran several iterations using different independent variables in order to identify those that were most correlated with the lead time. Once the regression model was working reasonably well, we used it to estimate lead times for a sample of 100 SKUs. The advantage of an accurate estimation of the lead time is to give the sponsor company more predictability in their supply chain. This predictability will allow them to take advantage of opportunities to handle certain SKUs with more cost-effective supply chains. For example, if they know a product will have a three month lead time, they do not need to import it immediately and hold it in their distribution center. They can hold the item at origin and use more economical storage and transportation options. In order to demonstrate this value, we use our model to delay the export of shipment of SKUs that we predict will have an abnormally long lead time. We can simulate the resulting time spent at DCs and compare it to the actual time spent at DCs to predict potential savings. 20 4 Data Analysis and Results We analyzed 2013 purchase order data for shipments from China to US. The analysis consisted of four stages: 1) initial statistical data description; 2) hypothesis; 3) model construction; 4) tests and results. In the first stage, we looked at transit time statistics for shipments from China to US, including averages, standard deviations, and distributions. Then, we broke down the total lead time into several components in order to understand how each part behaved and contributed to the total lead time, as shown in Figure 3. 21 i* fD P0 Create - Cargo Received PO Creat Date Book Date Confirmation Date Month PO CreateI Sail Date Actual Receipt Date lConsolidation Date CY Book Date Con Can Date CFS Actual Receipt Date Consolidation Date Ui| PO Create Datel Itrt Shp Date PROCESS DETAILS ORDER DETAILS ecpected at ETD Load Port Origin Service PO CreateDC Receipt kI Estimated Arrival SaIlate -, I 1 Estimate Arrival Deconsolidator Arrival Delivery BookedHeader Date . ....... .... .. .... Destination Service I First KeyL Date - Last Keytrol Date DC Receipt I I Deconsollor Date - Last Header Place of Arrive Decon Discharge Port Store Ready EFA PO Create Date - Header Date Actual Weight Actual Measurement Actual Packages Actual Quantity Vndor The purpose of this study was to find out how different supply chain strategies can be applied in order to reduce the dwell time at destination DCs and the relevant logistics costs. In order to do this, the second stage of our analysis focused on hypothesizing on what attributes on a given purchase order may help to explain why the dwell time at DCs behaved differently. Data visualization is used to help to build our initial hypothesis. Based on the visualization result, certain attributes are selected to be included in our hypothesis that can help to explain the variation of the dwell time at destination DCs. In the model construction stage, regression models are used to verify the correlation between the ten attributes and dwell time at DCs. Specifically, two types of regression models are used to carry out the analysis: multi-factor based regression model and ordered-probit regression model. Finally, in the result and test stage, out-of-sample data are used to test the accuracy of the regression models' predictions, and simulation is further used to test the effectiveness of the segmentation strategy based on the regression results. 4.1 Initial Data Statistical Description One year's Purchase Order data for shipments from China are given to perform the analysis. There are total 298,711 specific order lines. Unlike traditional segmentation approach, which typically used attributes such as profit margin, demand volume, demand volatility, value density etc., the challenge for our study is to segment the SKUs using only the attributes on the Purchase Orders. Figure 4 is the list of the attributes and their descriptions on a typical Purchase Order. 23 Dat Element Element Description If N then only TJX po data; If Y then Damco shipment Data; APLL older PO Data PO Shipped Division PONumber ) Oracle Agent Num Date UP (M/Y Po Create Date Strt Shp Date PO Merchandise Type Pretkt ind Prepk Vendor Numn Po Page Num Po Line Num Nesting Code Con Can 1Dte ind Style Ladder into Sku Indicator Style of set Indicates if Item Is Set Indicator Arrive Decon 1st Keytrol Last Header Deconsoliator Date - DC Receipt PO Create - DC Receipt First Keytrol Date - Last Keytrol Date PO Create Date - Header Date PO Create -Cargo Received PO Create -Sail Date Sail date -Estimated Arrival Estimated Arrival - Deconsolidator Arrival PO Create Date - Book Date Booked -DC Receipt -Header Date Booked Equipment Number Book Date Confirmation date Actual Receipt Date Create of goods received origin origin arrived of arrival at umber goods is of Consolidation Date ETD ETA Carrier Service Destination Load Port Country Load Port Discharge Port Country Discharge Port (Damco) Place of Delivery Ordered Quantity Actual Quantity Actual Packages Actual Weight Actual Measurement Origin Division Number PO Number PO Department Number TiX TJX Agent Number PO Cancel date at freight forwarder PO Plan Month (expected at DC) Date po was entered mainframe Earliest date Vendor can deliever the goods to consolidator Po Type Preticket Indicator PrePak Vendlor Number Page on Po for style number Line on PO for style number Code for pack together From ACT part a Date the goods arrived atthe deconsolnoator (US) First Date of cargo receipt at DC Date that goods were worked at the DC Timing: Arrival at Deconsolidator -Arrival at DC Timing: Po Create date - Arrival at DC Timing: First Date cargo receipt at DC - Date that goods were worked at the DC date - Date that goods were worked at the DC Timing: Po Timing: Po Create date - date were at Timing: Po Createe dat e goods sailed from Timing: Transit time for goods on the water port - Date goods at econsolidator Timing: Date Timing: PO Create date - Vendor booking date Timing: Vendor booked date - Arrival DC were worked at the DC Timing: Vendor booked date - Date that Container N Vendor Booking Date Date Damco confirmed booking Actual cargo receive date date that Consolidation date cargo: CF the date the container stuffed, CY the cargo has been received at the po Estimated Vessel Departure Date Estimated Vessel Arrival Date ship factory Mode (Consolidated by forwarder Destination Mode (Port Move Door Move) Export Country ExportPort Country discharge Port Port Arrival Final Destination on BL Po Ordered Units PO Shipped Units PO Shipped Cartons POShippedWeight(kgs) Po Shipped Cubic meters # Dept Num AP Vendor (Lawson Service line Carrier Steam Origin of vs or the load) of Figure 4 - PO Attributes and Descriptions We first wanted to understand how the different timing attributes behaved across the whole data set and statistically measure the time spent in different stages of the supply chain process, along with the variation and distribution of this timing. Most importantly, we wanted to understand how dwell time at destination DCs varies across different Purchase Orders, since this is the attribute we try to explain and predict in our analysis. 4.1.1 Measuring Timing Attributes We first looked at the total time spent for moving shipments from China to US in the given data set. We used PO Create - Header Date to measure this time attribute. This measure is equal to 24 the number of days between when a PO is created and when it is finished being processed at a DC and ready for distribution to retail stores. PO Create Date - Header Date tMspwiofPO OatDt- HOdr Date One Variable Summary Mean Variance Std. Dev. Skewness Kurtosis Median Mean Abs. Dev. Mode Minimum Maximum Range Count m 117.00 34.14 112.00 34.00 630.00 3rd Quartile Interquartilo Range 56.00 Ist Quartile MO1I baw0 596.00 284497 35660800.00 95.00 151.00 Sum e g a m 125.35 186114 43.14 0.8370 4.0091 g g Figure 5 - PO Create Date to Header Date Statistics (in days) As it is shown in Figure 5, the company spent an average of 125 days moving shipments from China to US, with a median of 117, standard deviation of 43 days, minimum of 34 days, and maximum of 630 days. So, its distribution is skewed to the right with a skewness of 0.837 and a kurtosis of 4. We then broke down this total lead time into several sub-components and looked at how they behaved statistically, as shown in Figure 6. Because our study will be used to help the company determine different supply chain strategies for moving future shipments, only the time attributes that cannot be determined at the time of decision making are of particular interest. Those time attributes are: 25 1) Estimated Time of Departure from origin port to Estimated Time of Arrival at destination port (ETD to ETA) 2) ETA to Arrival at Deconsolidator Warehouse 3) Deconsolidation Date to receipt at company Distribution Center (DC) 4) Date goods received at company DC to date goods were processed at DC Of these four attributes, the first three can be used as lead time distribution inputs in order to simulate the outcome of using different supply chain strategies. The fourth element partially captures the dwell time spent at destination distribution centers. Since destination dwell time (days between arrival at a DC and actual demand at retail store) is the key attribute we tried to explain and predict, our next focus is on how to define and measure this attribute appropriately. Figure 6 on the following page quantifies the different lead time components that were identified earlier. 26 II a I rilli I I I 5|-|3 I I ~I -j I~I I I 51 +; 2 II 'Ia 1;I I 2 I II 611111 AI 'III I Figure 6 - Lead Time Components with Statistics 27 4.1.2 Defining and Measuring Dwell Time Dwell time is the time difference between the goods received at DCs and the actual demand window. In order to correctly measure the dwell time, we first need to understand when the shipments were received at destination DCs and when they left. In the data set given, First Keytrol Date (as shown in Figure 1) represents the time the shipments are received. To measure the time shipments left a DC, our original thought was to use Last Header since this attribute represents the last date that the goods were worked at DC. However, during interviews with the company, we were told that the Last Header is not an accurate measure of the date a shipment leaves the DC for two reasons. First, Last Header is the date the last item was processed at DC. In many occasions, the majority of the inventory has already left the DC and any remaining pieces, no matter how small the quantity, would lead to an unrealistically late Last Header date for the entire PO. Second, the actual demand may happen before a shipment arrives at DC. In other words, the shipment was late and the Last Header date will be beyond the actual demand date. Therefore, in addition to the Purchase Order dataset, the company gave us the actual Ladder Plan dates to represent the actual demand window. The dwell time at DC is then equal to the number of days between the First Keytrol Date and the Ladder Plan Date. If the result is negative, it means the shipment was delayed and failed to fulfill the demand on time. 28 First Key to LP One Variable Summary Updated LP Date Histogram of First Key to LP/ Updated LP Variance 1.42 890.26 20OW Std. Dev. 29.84 16M Skewness Mean Abs. Dev. 5.3008 46.7737 -2.00 15.85 Mode -4.00 Minimum -108.00 Maximum 399.00 507.00 Mean Kurtosis Median "ane Date 33O 1 M LZM 0COO WW Figure 7 - Dwell Time Statistics (in days) As shown in Figure 7, the historical dwell time averages about I to 2 days with a standard deviation of almost 30 days, which indicates the historical dwell time is very volatile. Also, with the median of -2 days, more than 50% of shipments failed to arrive before the actual demand window. In the next section, we will build our hypothesis regarding which attributes can help to explain this dwell time variation. 4.2 Hypothesis There are many factors or attributes that may have some level of explanatory power of the dwell time variation. In this study we are limited to the attributes on the purchase orders, and only the attributes that are available before moving shipments from their port of origin can be used. Attributes that are created after that are hind-sighted information and therefore cannot be used to predict dwell time. The attributes that can be used for this study are listed below: 1) Division - The sponsor company operates multiple retail chains. Each one is a considered a division in this data. 29 2) Department - Each item has a department number that represents the category of item (i.e. shoes, menswear, etc.) 3) Vendor - Each supplier is assigned a unique ID. 4) Agent - Some products are sourced through agents, which are also assigned unique IDs. 5) Start ship date - The date that suppliers are able to begin shipping product. 6) Con cancel date - The deadline for suppliers to arrange shipment of product. 7) Merchandise Type - Proprietary 8) Ladder Plan Month - The month that an item is expected to be for sale in retail stores. 9) Place of Delivery - Destination location 10) Loading Port - Origin port 11) Style Ordered Quantity 12) Actual Quantity 13) Actual Packages 14) Actual Weight 15) Actual Measurement 16) Pre-ticketed - Whether or not an item has a price tag at the time it leaves the supplier facility. 17) Pre-packed - Whether items are packaged for store delivery by the supplier, or the company must repack them at their distribution center. 18) Time between PO create and Cargo Receipt at origin 19) Time between PO create and Start to Ship 30 .~ ..... .... ... ... ......... .. We then used data visualization to quickly identify which of those attributes can help to explain the dwell time variation. Sheet 1 Dept Num 101 102 400 103 109 350 112 113 C 300 114 115 C 110 *120 S124 geC 200 150 *120 128 U129 131 *133 0 235 CC...4. 50 100 236 6 238 *1A 150- 3500 2350 130 *362 *363 *366 PO Ordered Quantity -100 0 50 100 150 200 PO to Cargo Received 250 300 350 4130 1000 147000 Figure 8 - A scatterplot showing the days between PO Create and Cargo Receipt on Y-axis, and Destination Dwell Time on x-axis. Different colors identify different departments, and size of circle represents order quantity. For example, in Figure 8, we plotted the number of days cargo spent at the company DC (DC receipt to ladder plan date) against the number of days between the date the PO was created and the date the cargo was received at origin, based on Department Number and Ordered Quantity attributes. Visually, we can quickly identify that it seems that as PO to Cargo Received time increases, the dwell time will decrease. In Figures 9 and 10, we see that Dept 236 has much more dwell time variation when compared to Dept 346. 31 Sheet Dept Nun, 1 4W0 350 e., 150 : . a - 250 I * ON * IM PO 100 Ordered Quantity S500DO 50 100 200 PO to Cargo 250 300 350 1 OD0O 147000 400 Recoied Figure 9 - Figure 8 has been modified to show a department with more variability than normal. Dept Num Sheet I 400 350 300 250 200 150 100 ME 50 ii PO Ordered Quantity ) 50000 50 100 20 150 PO to Cargo Re ~e 250 300 350 400 1001 147000 Figure 10 - Figure 8 has been modified again to show a department with stable timing. 32 By using this data visualization approach, we narrowed down the original attributes to create the following list: 1) Department (Binary Variable) 2) Merchandise Type (Binary Variable) 3) Ladder Plan Month (Binary and Numerical) 4) Place of Delivery (Binary Variable) 5) Dvision (Binary Variable) 6) Pre-ticketed (Binary Variable) 7) Pre-packed (Binary Variable) 8) Time between Cargo Receipt and LP month (Numerical Variable) 9) Time between PO create and LP month (Numerical Variable) Our hypothesis is that these nine attributes can help explain the dwell time variation. In order to test our hypothesis and understand the numerical relationship between these nine attributes and dwell time, a numerical model must be constructed. In the next section, we discuss our model construction and results. 4.3 Model Construction In the previous section, we hypothesized that there are nine attributes that may help to explain the dwell time variation. At this stage, we use various multifactor regression models to test this hypothesis and determine the numerical correlation between the variables and the dwell time. The company is particularly interested in segmenting future shipments into the following groups so they can use different supply chain speed to improve on-time performance: 1) Late shipment: dwell time is less than negative 21 days 33 2) On time shipment: dwell time is between negative 21 days and 0 3) Early shipment: dwell time is greater than 0. The remainder of this section documents the development of an ordered-probit regression model and a neural network regression model that proved to be most effective in identifying correlation and predicting dwell time. We then provide an example of how a regression model can be used to form a segmentation strategy for the purchase orders based on the predicted dwell time and simulate the impact on on-time performance. 4.3.1 Ordered-Probit Method Since our objective is to determine a method to segment future purchase orders into three groups, we decided to use ordered-probit method to construct our model instead of using simple linear regression method. We assumed this model would be more accurate and reliable because it predicts the probability of dwell time falling in a given range rather than predicting an exact number of days as in a linear regression. Unlike standard linear regression models where the dependent variable needs to be numerical, an ordered-probit method is a technique used to regress categorical values against the chosen attributes. Instead of a predicted value, the output of a probit model is the likelihood of the occurrence of each category given a set of inputs. A tyical ordered probit model can be described as follows (Chris Brooks, 2008): =i xi /3 + EL Where ei are independent and identically distributed random variables, xi is the independent variable matrix, and 0 is the coefficient matrix. Then, yi is determined from yi* according to the rules below: 34 if y* :5 y1 ify1<yt<y2 if yi* > Y3 1 yi = 2 (3 Then, the probabilities of observing each value y are given by: P(yj = 11x,f, y) = F(y1 - xi' f) F(y 1 - x i) P(y = 21xi,f, y) = F(y2 - xi' P(y = 31x,fl,y) = 1 - F(y2 - xi #) The yi, the threshold values, and f , the coefficient matrix will then be estimated according to the natural log likelihood function: Maximize: L( 53, y) ln(P(y1 = = jIxi, fl, y)) (yi = j) i=1 j=1 is a logical value. If y = j is true it will take a value of 1, In the above formula, 1(y = j) otherwise it will be 0. F( y x' f - ) is the cumulative probability function of the error terms, which is assumed to be normally distributed. To build an ordered-probit regression model, the dependent variable - dwell timein this case - is translated into three categorical ranges. Before running the model, we still have to transform the original dwell time values into the categorical values, which are detailed as follows: " Categorical value: 1, for dwell time <=-21 days * Categorical value: 2, for dwell time >-21 days and <= 0 day o Categorical value: 3, for dwell time > 0 day Next, we constructed the model in EViews, which is an econometrics software. In choosing the maximizing optimization method, the default Quadratic hill climbing method is used, the model output is summarized in Figure 11: 35 Pseudo R-squared Schwarz criterion Hannan-Quinn criter. 0.364280 1.247907 1.237530 Akaike info criterion Log likelihood Restr. log likelihood 1.232705 -23118.32 -36365.56 LR statistic 26494.47 Avg. log likelihood -0.614571 Prob(LR statistic) 0.000000 Figure 11 - Ordered Probit Model Output Summary The correlation coefficient between the dependent and independent variables as well as the significance level can be found in Appendix I. Since this is an ordered probit model, those coefficients cannot be directly interpreted as how much the dependent variable will change by changing one unit of the independent variables. Usually, marginal effects are calculated and used to measure how much the probability of the dependent variables will change given a unit change of the independent variables. Nevertheless, those coefficients can still be used to get a sense of the relative strength in correlation with the dependent variables among the independent variables. Specifically, the higher the coefficient, the higher is the chance that the dependent variable will fall into category 3. Regarding the model fitness, as indicated in Figure 11, the Pseudo R-squared value is 36%. However, since this is an ordered-probit model, this R-square cannot be directly interpreted as a measurement of goodness-of-fit. To better measure this model's predictability, the model outcome should be compared to the constant probability in the original dataset. This is to measure how the model performs in prediction relative to always choosing the category that has the highest historical number of occurrence, which in our case will be category 2 that has 45% occurrence in the past. Figure 12 shows the comparison of the model ability to predict the dwell time: 36 Prediction Evaluation for Ordered Specification Equation: EQ02PROBIT Date: 04/20/14 Time: 11:15 Estimated Equation Dep. Value Obs. Correct Incorrect % Correct % Incorrect 1 4233 1438 2795 33.971 66.029 2 17034 14570 2464 85.535 14.465 3 16350 13543 2807 82.832 17.168 Total 37617 29551 8066 78.558 21.442 Constant Probability Spec. Dep. Value 1 2 3 Total Obs. 4233 17034 16350 37617 Correct Incorrect 4233 0 16350 20583 0 17034 0 17034 % Correct % Incorrect 0.000 100.000 0.000 45.283 100.000 0.000 100.000 54.717 Gain over Constant Prob. Spec. Dep. Value Obs. Constant Equation % Incorrect % Incorrect Total Gain* Pct. Gain** 33.971 1 4233 66.029 100.000 33.971 2 17034 14.465 0.000 -14.465 NA 3 16350 17.168 100.000 82.832 82.832 Total 37617 21.442 54.717 33.275 60.812 Figure 12 - Ordered-Probit Prediction Evaluation Overall, this model achieves about 78% prediction accuracy. When compared to the constant probability, the model improves prediction performance by 33%. However, the limitation of this model is that it can only predict categories two and three with high accuracy. Its accuracy for category one is only 33%. Moreover, since this model incorporates many binary variables, it will frequently encounter multicollinearity problems, that were fixed in the final models, but could resurface when running the model over new datasets. 37 4.3.2 Neural Network Model Due to the limitations of the ordered-probit model, we used a neural network regression model in an effort to improve prediction accuracy, goodness-of-fit, and application feasibility. 4.3.2.1 Neural Network Model Mathematical Equations and Algorithm Neural network is usually referred to the various mathematical models that imitate human brain functions (Huangjin Tang, Kay Chen Tan and Zhang Yi, 2007). It has wide applications in areas such as artificial intelligence, information processing, engineering, and finance for pattern recognition, function approximation, forecasting purposes. In statistical forecasting, unlike traditional regression model which relies on modeler's ability in guessing the causal-and-effect function between the inputs and the outputs, neural network method provides a way directly learn the internal relationship between the variables in a system (Ke-Lin Du & M.N.S Swamy, 2014). This feature is very helpful when the underlying relationship between the inputs and outputs in a system cannot be very well estimated. The most widely used neural network model in statistical forecasting is feed-forward networks. A typical feed-forward network structure consists of a layer of inputs, a single or multiple hidden layers and an output layer. A typical feed-forward neural network model's mathematical equations can be described as follows: 1. Input variables and non-linear transform function: =(S j) Xj 0 d(l-1) ) (- (1w 1= 38 6(S) es - e-s es + e-s = Where: 1 1 L (1 stands for which layer in the model, L is the largest layer in the model) 0 i 1 j < d(') (i stands for which output in the model, d is the number of outputs) d(-')(i stands for which input in the model, d is the number of outputs) x is the ith output at layer (1 - 1)and the ith input at layer I x!is the jth output at layer 1 w!- is weight 0(S)is a tansigmoidnonlineartransform function for the input variable 2. Output and error term: The final output of the model is a linear output. The error term, which is the difference between the model output and actual value, is often measured by mean squared error. Based on the equations for the input variables, the error term is essentially a function of w!b . To minimize the error term, w!b can be determined by: de(w) Ve(w): de(w) a (1 awi de(w) sl __S_ aw(1) Since: 39 x. , so only 8w (w) needs to be calculated. S~l 3. Back propagation: To find the optimal weight, neural network regression models use a method known as back propagation. The error terms are run through the network layers in reverse order, then a gradient descent method is used to determine the optimal weight that minimizes the error term (Ke-Lin Du & M.N.S Swamy, 2014). Its mathematical equation usually can be described as follows: 80) =e(w) as~l Starting from the last layer: d(l) -ae(w) 81 d =1 i j=1 a _ x> e(11)l =x i i d(1) _ i = 8 x1w0f( x 6' (SO'~) j=1 4. Algorithm: A typical back propagation neural network algorithm follows the following steps: a) Randomly initialize w (1 b) Forward compute all x! in the net ) c) Back propagate all 6(1 d) Update wf e) Iteration f) Find the optimal wf that gives the lowest e(w) The neural network mathematical structure also has many derivatives which can use different non-linear transform functions, error term measurements, layer structures and so on. In this study, the modelling is based on the back-propagation feed forward model described above. 40 More specifically, the neural network model used in this study has one input layer, one output layer, and 10 neurons. 4.3.2.2 Model Building in Matlab To build, run, and test the neural network model, we used Matlab technical computing software. 3,000 purchase orders are first used to build the model. A rolling forecast mechanism is designed to roll the model over the entire 37,617 purchase orders in order to fully test the accuracy and robustness of the model. Parameters such as model size, variables to be included, and which part of the dataset to use to build the model and forecast, can be entered as the script starts to run. Because the sponsor company would only be able to use data from POs that have been delivered, we include a gap of 3,000 purchase orders in the initial model to represent the time gap needed to update the model. The size of this gap is another parameter that can be controlled. The detailed Matlab script for our model is presented in Appendix II. Also, in choosing the input variables, we narrowed down number of input variables from the nine variables used in the ordered probit model to five variables, which are Merchandise Type, Pre-ticked, Pre-packed, Division, and Time between Cargo Receipt and LP Month. The reason is that in building our neural network model, we found that adding the Department, Time between PO creation and LP month, and Place of Delivery variables didn't help to improve the performance of the model, and adding them into the model will simply complicate the application of the model, since it will ask users to provide more inputs than necessary. We use this script and the first 3,000 purchase orders in the dataset to build, train, and validate a neural network model. Figures 13, 14 and 15 show the resulting model construction. 41 Best Validation Performance is 106.2722 at epoch 7 10 - 10, 0 h Train Validation Test Best 3 So 10 _ 102 10 10 0 _k 0 4 2 6 8 12 10 13 Epochs Figure 13 - Model Training and Validation Error Histogram with 20 Bins 700 Training Validation Test Zero Error 600 500 0 500- 400 300 200 100 C 9 Errors = Targets - Outputs Figure 14 - Error Distribution Histogram 42 Validation: R=0.9243 Training: R=0.96247 t 250 d a + o 250 Data Fit Y= T -200 CP Data Fit ---- Y = T 200 150 150 100 D 0 100 50 50 0. o05 100 0 + 0 250 C4 200 150 All: R=0.95947 250 0' 200 LM 150 150 100 100 0 50 100 Test: R=0.965 Data Fit Y =T - 50 Target / 0 0 200 Target 50 0 200 250 0 Data -- -Fit Fit 0' 0 0 o -5o -50 0 100 Target 200 0 100 200 Target Figure 15 - Regress Predicted Value on Actual (In-sample test) As shown in figure 13, after training and validating, the model found the minimized MSE, which is the mean squared error between the actual value and predicted value, is at 106. Figure 14 shows and compares the error term distributions, which is the difference between the predicted value and actual value, during model training period, validating and train periods. The allocation of the sample data for training, validating and testing is 70%, 15% and 15% respectively. In figure 15, the predicted value, which is the output of the model, is regressed on the actual value, which is the target of the model. The dashed line represents the ideal situation where the output value is exactly equal to the target value and the R squared value is 1, while the solid line represents the actual fitness. Overall this model has very good fitness as indicated by the high R 43 squared value. Next the model is used to predict the dwell time for 100 out-of-sample purchase orders. Figures 16 and 17 show the result: : R=0.99273 o 200 - Data Fit -- '0 Cl! 150 'D 100 50 0 0 0 50 100 150 200 Target Figure 16 - Regress Predicted Value on Actual (Out-of-Sample Prediction for 100 POs) I I I I I I I I I - - - I I I I I - I I I - - - I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I I 60 SI 1 10 20 30 4 I 0 0 8 100 Figure 17 - Predicted Result VS Actual Dwell Time (Out-of-Sample Prediction for 100 POs) In this out-of-sample prediction, the model achieves fitness of 99% with a constant bias of 4.2 days. In the following section, more tests are performed to evaluate this model's accuracy and robustness. 44 4.4 Tests and Results In this section, both the ordered-probit model and the neural network model are tested and evaluated in terms of their predictive abilities, fitness, and feasibility in application. 4.4.1 Ordered-Probit Model Due to the multicollinearity problem caused by the use of binary variables, testing for model by choosing different sample size becomes difficult. The chosen binary variables that work over the first 10,000 samples may have multicollinearity problems when the model is built from the next 10,000 samples, and the binary variables that cause this problem will have to be deleted from the model, which will make the test results become incomparable. Therefore, we kept 36,617 purchase orders as our sample size, and run the model to predict the dwell time for the remaining 1,000 purchase orders in the dataset. Below is the result: Table 2 - Ordered Probit Model - Test Result for 1,000 POs Actual Predicted correctly Accuracy Category 1 111 25 22.52% Category 2 Category 3 467 422 128 290 27.41% 68.72% Total 1000 443 44.30% S J The prediction accuracy for this test is 44.3%, which indicates that the out-of-sample prediction power of the ordered probit model is not accurate enough for the company to make business decisions. One approach often used to improve the forecasting performance is to reduce the number of categories to be predicted. however, in our study, when we used logit model to predict one category, the forecast accuracy was even worse with the accuracy lower than 10%. 45 4.4.2 Neural Network Model Unlike the ordered-probit model, using binary variables in a neural network model does not cause any problems in running and testing the model. Therefore, the neural network model can be tested over any portion of the given dataset. There are two ways to do this: 1) Build a single model to predict the entire dataset Under this method, the first 3,000 purchase orders are selected to build the model. This model is used to predict the remaining purchase orders. 2) Build initial model then update it on a rolling basis to predict the rest of the dataset Under this method, the first 3,000 purchase orders are selected to build the model. This model is then updated using the rolling window method to predict the rest of the dataset. For example, if it is set to predict 500 purchase orders per run, it will use the first 3,000 POs to build the model and predict the dwell time for the POs from 6,001 to 6,500 (including the timing gap size of 3,000 POs). For the next run, POs from 1 to 500 will be excluded from the model and POs from 3,001 to 3,500 will now be included in the model used to predict dwell time for POs from 6,501 to 7,000. In either case, it is important to know how far this model can predict into the future while maintaining an acceptable accuracy level and what sample size results in the most accurate predictions. In other words, how frequent the model should be run in application to achieve the highest accuracy. The following table summarizes the test results under the two approaches: 46 Table 3 - Test Results without updating the model (model size=3,000, gap=3,000) How far into the future 100 POs 500 POs 1000 POs 5000 POs 10000 POs 30000 POs Fitness 99.27% 98.23% 98.30% 97% 96.50% 92% Coefficient 1 1.1 1.1 1.1 0.98 0.89 Bias (days) 4.2 4.9 5.9 2.9 1.8 0.69 Table 4 - Test Results with updating the model (model size=3,000, gap=3,000) Number of POs updated per run 100 POs 500 POs 1000 POs 1500 POs 3000 POs Fitness 86% 86% 85.80% 83.70% 82% Coefficient 0.92 0.87 0.88 0.93 0.82 Bias (days) 1.5 1.5 1.4 1.5 1.4 As indicated in Table 3, when tested under approach one the model achieves at least 92% fitness. The coefficient , which measures the slope of the linear relation between the actual value and predicted value, with one as the perfect forecast performance, averages at one and the bias is between one and six days. When tested under approach two, the model achieves at least 82% fitness, the coefficients vary between 0.82 and 0.92, and bias is 1.5 days. Based on these results, two observations can be generalized: 1) Prediction accuracy will drop as the time horizon is extended 2) The rolling approach results in less fitness accuracy but is less biased In the initial tests, the model size is fixed at 3,000 POs. The next step was to test the sensitivity of the model performance relative to the model size. Figure 21 indicates that increasing the 47 model size generally improves performance. Additionally, once the model size crosses a threshold the performance becomes less sensitive to changes in model size. Fitness Coefficient 0.34 0.59 500 Model Size 700 0.94 Bias 3 1.1 2.5 1000 0.98 1 5.9 2000 3000 4000 0.97 0.98 1 1.1 -10 5.9 0.986 5000 6000 7000 0.97 0.95 0.97 1 0.97 0.93 0.99 2.21 3.5 2.5 10000 0.965 0.92 2 1.6 02 04 02 0 0 r 4400 M 60-tj 8000I Figure 18 - Performance Sensitivity Relative to Change of Model Size (Number 1 X IL IP )X) of prediction=1,000 POs) 4.5 Potential Application of Model (Proof of Concept) Because the neural network model showed a clear advantage in prediction accuracy, we will not explore the potential use of the ordered-probit model. The predictions from the neural network model only solve part of the actual problem. How to segment the POs based on the predicted dwell time and how to apply different supply chain speed to improve on-time performance are the next steps to be taken for this model to be useful. The following example demonstrates how this model can be implemented by the company. First, 3,000 POs are chosen to build, train, and validate the model. As mentioned previously, the next 3,000 POs are excluded from this simulation to represent POs that have not been delivered and therefore do not have complete data. The detailed parameters of the model are summarized in Table 5: 48 Table 5 - Model Parameters Used in the Application Example Model Size Gap Size Input Variables Number of POs to be predicted 3,000 POs 3,000 POs Merchandise type, Preticked, Prepacked, Division, Cargo Received, LP Month 100 POS The output of the model will be the predicted dwell time for the next 100 POs. Then, the 100 POs will be segmented into the three predefined categories: * * * Categorical value: 1, for dwell time <=-21 days Categorical value: 2, for dwell time >-21 days and <= 0 day Categorical value: 3, for dwell time > 0 day Table 6 demonstrates how this part of the process would look in application: Table 6 - Segment Predicted Dwell Time into Predefined Categories PO 6001 6002 6003 6004 6005 6006 600 6008 6009 6010 6011 6012 6013 601 6015 6016 6017 601 6019 6020 6021 6022 6023 6024 6025 Predicted 5 5 133 5 5 163 5 5 -19 4 -19 3 3 -21 -15 13 13 13 -17 -21 3 3 -21 -21 -19 Actual 13 13 125 13 13 167 13 13 -15 13 -15 17 17 -14 -15 13 13 13 -15 -21 7 6 -11 -11 -13 Predicted Category Actual Category 3 3 3 3 3 3 3 3 3 3 3 3 2 3 2 3 3 2 2 3 3 3 2 1 3 3 2 2 2 3 3 3 3 2 3 2 3 3 2 2 3 3 3 2 1 3 3 2 2 2 Accuracy 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 The next step is to determine how the company can adjust supply chain speeds based on the predicted segments in order to improve on-time performance and/or reduce costs. For demonstration purpose, it is assumed that the company will adopt the following strategy as shown in Table 7: 49 Table 7 - Example Strategies for Supply Chain Segments Shipment Definition Category 1 Late shipment Category 2 On time shipment Category 3 Early shipment Supply Chain Speed TransitTime Average-l1 days Fast Average Normal Slow Average+predicted dwell time In Table 7, the average transit time refers to the historical average Cargo Received Date to First Keytrol Date, which is 32 days with its distribution shown in Figure 13. Sheet 3 Figure 19 - Histogram of Actual Cargo Receipt to First Keytrol Next, the transit time for those 100 POs (number of days between cargo receipt at origin and delivery to TJX at destination) is changed according to the strategy in Table 7. Then, a new dwell time for those 100 POs is calculated based on this change. 50 Actual Dwell Time 1 2C X 45 C 0 ac Figure 20 - Histogram of Actual Dwell Time for 100 POs Improved Dwell Time w Figure 21 - %fP aa C 60 2C 5C C 90 C Histogram of Improved Dwell Time for 100 POs As Figures 26 and 27 indicate, before the improvement, 6% of POs arrived late and 58% of the POs arrived too early. Among the POs that arrived early, 93% of them arrived at least 20 days earlier than needed. If the company were to use the simple segmentation strategy determined by the model, there would have been no late POs and 60% of them would have arrived exactly on time. The remaining 40% of the POs would still arrive early, but would be early by fewer than 10 days. 5 Conclusion Based on the tests of the models, we concluded that the neural network model achieves much greater prediction accuracy than the ordered-probit model in this study. Additionally, the neural network model can be easily updated to incorporate new information. This is a very useful 51 feature in actual application, especially for a company with seasonality. The ordered-probit model cannot be updated as easily due to muliticollinearity problems caused by binary variables. Therefore, the neural network model can be used by the company to predict the dwell time and segment the POs based on the result. However, one limitation of the neural network model is that the w! is initialized randomly when running the model, this stochastic process can sometimes make the model generate high errors. In this study, the model typically has a MSE between 30 and 100. When the MSE is above 200, the model prediction accuracy tends to become unstable. One way to correct this problem is to run and train the model several times until the error term is down to a desired level. The simulated application of this model shows that it is feasible to implement a segmentation strategy based on the output of the neural network model and on-time performance can be improved significantly. The six percent of POs that were delivered late could be expedited and potentially increase sales. The shipments that arrive slightly early (less than a month) could be shipped using slower and more economical transportation options. In the simulation, this accounted for about half of the POs which means the potential savings are very significant. The shipments that arrive very early are perhaps the highest area of opportunity. These purchases can be acquired later from suppliers, or could be stored at origin. This will reduce holding costs at destination DCs and free up space in their network that is used to store early shipments. This is a great benefit to the sponsor company as it will allow them to increase volume without making costly expansions their distribution network. To maximize the performance of the model, we suggest the company capture additional data that could be relevant to supply chain timing when creating POs. For instance, when touring distribution centers we observed a lot of volatility in the amount of time different SKUs spend in 52 the distribution center before they are ready to be shipped to stores. This required processing time is currently invisible to the company, but could be recorded and used to improve adjust transit times for future orders. This information and additional data such as historical on-time performance of different suppliers and buyers, or a metric that buyers can use to communicate the urgency of POs, could be used to improve future versions of the model. Because there is always uncertainty when using a predictive model, these metrics could also help the company decide when they should buffer the model output. 53 Appendix I - Ordered Probit Model Output: Variable Coefficient Std. Error z-Statistic Prob. PRETICKTY PREPKTY JAN MAR MAY JUN AUG SEP OCT NOV TYPE_1 TYPE_2 TYPE_3 TYPE_6 -0.074357 0.085350 -0.059730 -0.295996 0.380402 0.004634 0.063681 -0.349777 0.233088 0.078296 -6.456124 -6.356646 1.364158 -6.202059 0.044614 0.023546 0.029739 0.026030 0.029859 0.028680 0.028568 0.025396 0.026165 0.022754 2389.213 2389.213 2404232. 2389.213 -1.666660 3.624768 -2.008460 -11.37137 12.73993 0.161569 2.229140 -13.77313 8.908471 3.440889 -0.002702 -0.002661 5.67E-07 -0.002596 0.0956 0.0003 0.0446 0.0000 0.0000 0.8716 0.0258 0.0000 0.0000 0.0006 0.9978 0.9979 1.0000 0.9979 TYPE_7 -6.278583 -6.519581 2389.213 -0.002628 0.9979 TYPE_8 2389.213 -0.002729 0.9978 POD_881 -0.627139 0.188543 -3.326244 0.0009 POD_884 -1.701864 0.450726 -3.775824 0.0002 POD_885 POD_955 POD_964 POD_970 NUM102 NUM103 -1.602656 -2.004570 -1.317218 -1.628698 -0.026699 -0.186398 0.257253 0.062149 0.024036 0.033680 0.058170 0.065245 -6.229885 -32.25430 -54.80076 -48.35846 -0.458972 -2.856896 0.0000 0.0000 0.0000 0.0000 0.6463 0.0043 NUM111 -0.130785 0.140679 -0.929671 0.3525 NUM112 NUM116 NUM120 1.487328 -12.63921 -0.720489 0.588088 1649925. 0.138239 2.529090 -7.66E-06 -5.211902 0.0114 1.0000 0.0000 NUM124 -0.326453 0.083075 -3.929623 0.0001 NUM128 NUM129 NUM236 NUM346 NUM347 NUM348 NUM349 NUM356 NUM358 NUM359 NUM360 NUM362 NUM363 NUM366 NUM450 NUM451 NUM452 NUM453 NUM455 -0.791264 0.032269 0.181421 0.037849 -0.317279 3.847498 -0.283656 0.287244 0.067265 0.020737 -0.021652 -0.064302 0.134225 0.258893 -0.322776 -0.359345 -0.457402 -0.367397 -0.024978 0.133675 0.065884 0.048861 0.039086 0.039231 1744173. 0.054707 0.041984 0.159771 0.040866 0.049639 0.055149 0.055393 0.049101 0.044953 0.051123 0.037862 0.048793 0.101328 -5.919305 0.489777 3.713000 0.968342 -8.087480 2.21E-06 -5.185038 6.841804 0.421010 0.507434 -0.436180 -1.165955 2.423134 5.272661 -7.180389 -7.029051 -12.08081 -7.529682 -0.246511 0.0000 0.6243 0.0002 0.3329 0.0000 1.0000 0.0000 0.0000 0.6737 0.6119 0.6627 0.2436 0.0154 0.0000 0.0000 0.0000 0.0000 0.0000 0.8053 54 113.0017 -2.708509 0.0976 0.2060 0.0244 0.0012 0.3927 0.2182 0.1039 0.0973 0.0000 0.0007 0.1212 0.0001 0.1411 0.0000 0.0000 0.0000 0.0068 -0.002321 -0.001413 0.9981 0.9989 0.212175 0.128076 1.656635 0.131815 -1.264550 2.251281 NUM539 NUM541 -0.166687 0.095850 0.306472 -0.049565 -0.228738 0.116640 0.201158 0.566689 0.199860 0.292992 0.203503 0.093107 NUM643 -0.345179 PO_TO_LP CTOL DIVISION 1 0.001152 0.066607 -0.101180 NUM469 NUM472 NUM475 NUM479 NUM480 NUM481 NUM484 NUM485 NUM486 NUM489 NUM490 0.042576 0.094646 0.057992 0.185760 0.071732 3.238086 -0.854689 -1.231368 1.626063 0.121329 1.657953 0.104034 0.058857 0.189039 0.052181 0.063268 0.064372 0.000258 0.000589 0.037356 5.447132 3.395705 1.549900 3.899922 1.471636 -5.362276 4.467529 Limit Points LIMIT_2:C(66) LIMIT_3:C(67) -5.545330 -3.376751 2389.213 2389.213 1.232705 0.364280 Akaike info criterion Schwarz criterion 1.247907 Log likelihood -23118.32 Hannan-Quinn criter. LR statistic Prob(LR statistic) 1.237530 Restr. log likelihood Avg. log likelihood -0.614571 Pseudo R-squared 26494.47 -36365.56 0.000000 55 Appendix II - Matlab script for neural network model "Dwell Time Forecast start=input('Enter starting row number:'); ending=input('Enter ending row number:'); s=input('Enter size of the model:'); gap=input('Enter gap size:'); incr=input('Enter the number of POs to be predicted:'); sv=input ('Enter starting variable number:'); ev=input('Enter ending variable number:'); storedprediction=[]; storedactual=[]; for n=(start:incr:ending) in=inputmatrix(n:(n+s-1),sv:ev); output=outputmatrix(n:(n+s-1),l); inputs = in'; targets = output'; Create a Fitting Network hiddenLayerSize = 10; net = fitnet(hiddenLayerSize); Set up Division of Data for Training, net.divideParam.trainRatio = 70/100; net.divideParam.valRatio = 15/100; net.divideParam.testRatio = 15/100; Validation, Testing % Train the Network [net,tr] = train(net,inputs,targets); Test the Network outputs = net(inputs); errors = gsubtract(outputs,targets); performance = perform(net,targets,outputs); % Predict dwell time predictinput=inputmatrix( (n+s+gap) : (n+s+gap+incr-1) ,sv:ev); predictoutput=net(predictinput'); storedprediction=[storedprediction;predictoutput]; storedactual=[storedactual;outputmatrix( (n+s+gap) :(n+s+gap+incr-1) ,1)]; end storedprediction=storedprediction'; storedactual=storedactual'; result=storedprediction(:); actual=storedactual(:); [r, c]=size(result); figure plot(l:r,result,'r') hold on plot(1:r,actual) legend('result','actual') grid on figure plot(1:r,actual) plotregression(actual,result) errorscalar=actual-result; squarederror=errorscalar.^ 2 ; meansqurederror=sum(squarederror)/r; 56 References Brooks, C. (2007). Introductory Econometrics for Finance. Cambridge University Press. Du, K.-L., & Swamy, M. (2014). Neural Networks and Statistical Learning. New York: Springer. Fisher, M., Hammond, J., Obermeyer, W., & Raman, A. (1997). Configuring a Supply Chain to Reduce the Cost of Demand Uncertainty. Production and Operations Management, 211225. Lovell, A., Saw, R., & Stimson, J. (2005). Product value-density: managing diversity through supply chain segmentation. The InternationalJournalof Logistics, 142-158. Tang, H., Tan, K. C., & Yi, Z. (2007). Neural Networks: Computational Models and Applications. New York: Springer. 57