2014 execsummary gilliganjin
- 1. KEY INSIGHTS 1. A traditional SKU segmentation based on high level attributes such as volume, margin, and demand may not work for a company that is constantly changing their product mix and suppliers. 2. Analysis of historical data can help identify which product attributes might indicate the required supply chain speed for a given product. 3. When segmenting using high level attributes is not reliable, a mathematical model can be used to predict required lead times at the PO or SKU level. By Brad Gilligan and Huiping Jin Thesis Advisor: Edgar Blanco Summary: In this thesis, we determine how a global retailer can adjust their supply chain process for different SKUs in order to better meet demand and/or reduce costs. We do this by using purchase order data to predict which products require an expedited process and which can be managed in a more cost-effective manner. Because the product mix was too diverse and dynamic to use traditional metrics such as product value and demand patterns, we develop a regression model that the company can use to make supply chain decisions. Brad came to the SCM program with over three years of experience procuring logistics services as a Senior Procurement Analyst at Mattel. After graduating, he joined Coyote Logistics as a Supply Chain Strategy Manager. Huiping Jin received a MS Finance and an MBA. Prior to the SCM program, he worked as a logistics manager in manufacturing industry for over 10 years. Upon graduation, he will join Cummins as a supply chain manager. Introduction The cost of transportation, excess inventory, and lost sales can vary greatly between products. This is why it does not make sense for companies to use a one- size-fits-all supply chain for every product. Instead, companies should create multiple supply chain configurations that are tailored to the types of products they sell. This is commonly achieved using SKU segmentation designed to optimize the tradeoff between supply chain speed and cost. Our sponsor company asked us to perform a segmentation for them, but their unique business model created some challenges. Methods The sponsor company has suppliers and retail stores in multiple countries, but in order to make the scope of this project more manageable we focused only on items that are purchased in China for sale in the United States. This data alone included over thirty thousand Purchase Orders (POs), forty thousand unique SKUs, and one thousand different suppliers. Our approach involved thorough qualitative analysis and process mapping in order to understand the current supply chain from procurement to store delivery. We toured deconsolidation centers and distribution centers, and interviewed employees across the organization. After gaining this understanding, we moved on to analyzing historical shipment data. We used a combination of histograms, box-plots, and scatter-plots to illustrate the supply chain timing of products based on various attributes. These visuals helped us detect patterns and trends in supply chain SKU Segmentation for a Global Retailer
- 2. timing that could be related to PO data. We then built models to validate these relationships. Linear regression and ordered-probit regression models were able to assess the strength of relationships between PO data and supply chain timing. However, these models could not predict the required speed with much accuracy. We eventually developed a neural network model that achieved the desired predictive capability. Analysis Our research of the current supply chain process helped us understand the following key milestones: • PO Create Date – Buyer agrees to buy items from the supplier. • Cancel Date – Supplier is required to deliver the PO at origin, or the order can be cancelled. • Origin Receipt Date – Cargo is actually received at origin. • Destination Receipt Date – Cargo is received at destination distribution center. • Sale Date – Product is expected to be at retail store. The problem is that some products are received well in advance of the sale date while others are received without much time to get from China to a store in the US. Because the company cannot see which products may fit one of these descriptions, some products arrive at destination very early and take up space in distribution centers while others do not meet the planned sale date. The following histogram shows the number of days between the date a product arrived at the company’s distribution and the date it was needed in order to meet their sales plans, we referred to this amount of time as the destination DC dwell time. Figure 1: Histogram showing the number of POs that had a dwell time within the specified range (number of days) The large values on the right show the opportunity to utilize a slower more cost-effective supply chain process, and the negative values on the left show where products could have been expedited in order to meet sales plans. The objective of our model is to adjust these lead times so that more products will fall within the acceptable DC dwell time range of zero to seven days. Model After identifying data that appeared to be correlated with the dwell time, we first built a standard linear regression model. This model had very limited success in predicting dwell time. We then attempted to use an ordered-probit regression model, which we thought would be more effective because it does not predict an exact value but instead predicts the probability that a value will fall within a given range. This seemed appropriate for our application since we wanted to identify products that could be early, on- time, or late. Again, we did not have much success in predicting dwell times. However, these models did confirm the correlation of certain variables which we were able to use in our next and final model. Using the variables that showed a strong correlation with the dwell time, we developed a neural network model. The model consists of one hidden layer, 10 neurons and one output layer. The model starts by initializing a weight vector to calculate the initial predicted value. The predicted value is then compared to the actual value to calculate the error term. The weight vector is then adjusted by using the back-propagation method until the error is minimized. After training and validating the model, we tested the model performance using one year of actual data and found that our model has very robust forecasting performance and overall fitness that is consistently above 90%. The following figure demonstrates the model’s fitness and forecasting accuracy for predicting 100 PO’s dwell time: 0 5000 10000 15000 20000 -97.86 -57.30 -16.74 23.82 64.38 104.94 145.50 186.06 226.62 267.18 307.74 348.30 388.86 Frequency Histogram of DC Dwell Time
- 3. Figure 2: Graph showing the number of days of dwell time predicted by our model against the actual number of days Based on the predicted dwell time, we can then segment the purchase orders into three categories: • Late Shipments (dwell time <= -21 days) • On-time Shipment (dwell time >-21 days and <= 0 days) • Early Shipment (dwell time > 0 days) For each of the three categories, a different supply chain speed can be applied in order to optimize the on-time delivery performance. Conclusions Even with constantly changing suppliers and products, we were able to identify PO attributes that could be used to predict how much time our sponsor company has to get a product from origin to destination. In test simulations, these predictions were able to improve on-time performance from 36% to 60%. Additionally, no products arrived late and the products that did not arrive on time were less than ten days early. Figure 3: Histogram showing actual on-time performance and on-time performance that can be achieved using our model to adjust lead time The predictive capability of our model will allow the sponsor company to reduce transportation and holding costs because they will know when they have time to ship products slower or hold them at origin. They will also be aware of products that need to be expedited, allowing them to meet their sales plans more consistently and potentially increase total sales. Another advantage of the neural network model is that it can be updated automatically as new data is loaded into it. This means that the sponsor company can continue to use this model without repeating this entire process. 0 10 20 30 40 50 60 70 80 90 100 -50 0 50 100 150 200 250 Purchase Orders DwellTime result actual