Multi-Echelon Inventory Management for a Fresh Produce Retail Supply Chain

Transcription

1 Multi-Echelon Inventory Management for a Fresh Produce Retail Supply Chain By Thomas Hsien and Yogeshwar D. Suryawanshi Thesis Advisor: Dr. Amanda Schmitt Summary: Chiquita, a fresh produce distributor with a multi-echelon supply chain, wants to develop an optimal inventory policy that results in the lowest inventory costs while maintaining 95% service levels at each echelon. We develop a simulation model to determine the optimal inventory policy. The model s input parameters include inventory holding costs, shrinkage costs, lost sales costs, forecast accuracy and service levels. After using the model to determine the optimal inventory policy, we test the sensitivity of the multiechelon supply chain under the optimal policy with respect to forecast errors and transportation lead time. Thomas Hsien holds a B.S. in Business Administration and a B.A. in East Asian Language and Culture from University of Southern California. Prior to joining MLOG, he worked as a Reverse Logistics Process Manager for Sanyo. Thomas is joining Apple Inc. as a Global Supply Manager. Yogeshwar D. Suryawanshi holds a M.S. in Industrial Engineering from University of Illinois, Chicago and a B.Tech in Mechanical Engineering from National Institute of Technology, India. Prior to joining MLOG, he worked as a Consultant for Terra Technology. Yogeshwar is joining PricewaterhouseCoopers Supply Chain Advisory practice. KEY INSIGHTS 1. Lowering the target on-hand inventory levels at the distribution center and retail stores to 0.5 and 1.5 days respectively reduces total relevant cost for the system by 31% and maintains high service levels above the required 95%. 2. The multi-echelon supply chain under the optimal policy is sensitive to forecast errors. Total relevant cost for the system is more sensitive to forecast error at the distribution center, while item fill rates are more sensitive to forecast errors at the retail stores. 3. Chiquita should continue to minimize transportation lead time from the plant to the distribution center to maximize the product s available lifetime and reduce shrinkage. Introduction Perishability presents a challenging problem in inventory management for the fresh produce industry since it can lead to higher inventory costs and lower service levels. If a supply chain has multiple echelons, that further complicates the issue because companies have an added risk of not having the right amount of product at the right location at the right time. Additionally, the volatile demand of the fresh produce products at the retail level makes the issue even more challenging. Developing better order and inventory policies requires an understanding of the impacts of the product s limited lifetime, the interactions of multiple inventory locations, and the trade-off between the relevant costs and the customer service levels. We conducted our research on one of Chiquita s Fresh Express supply chains. We developed a simulation model using Arena simulation software to analyze the impact of perishability on total relevant

2 costs. Our research focused on determining the optimal inventory policy for the multi-echelon system considering inventory holding costs, shrinkage costs, lost sales costs, forecast accuracy and service levels. We tested the sensitivity of the system under the optimal policy with respect to forecast errors and transportation lead time. Due to confidentiality concerns, all the numbers used throughout the thesis are for illustrative purposes only and are not necessarily indicative of actual performance at Chiquita. Research Objective The objective of our research was to develop an inventory policy that minimizes costs for the system while maintaining high customer service levels at each echelon. The research answered the following three questions posed by Chiquita: 1) What are the parameters for optimal inventory management depending upon forecast accuracy, inventory carrying cost, product perishability, lost sales and inventory shrinkage costs? 2) What is the trade-off between service level and inventory costs? different virtual scenarios by changing the input parameters of the model. Additionally, a simulation model can help to keep track of the age of the inventory and monitor the interaction between the changes in each echelon of the supply chain. Problem Formulation/Model Construction In order to develop a simulation model that replicates the Fresh Express supply chain, we first investigated the standard practices, performance metrics, and challenges that exist in the supply chain. After we gained a good understanding of the key elements and data available in the Fresh Express supply chain, we mapped out the process flow of the supply chain and built a conceptual model to capture all the processes involved in this multi-echelon supply chain. Figure 1 shows the overview of the process flow for Chiquita s Fresh Express Supply Chain. Step 1 Step 2 Step 3 Step 4 Consumers Purchase Products at Retail Retail Stores Places Order at DC DC Ships Products to s s Receive Products 3) What is the impact of increased forecast accuracy on inventory related costs? Research Scope DC Places Order to Production Plant DC Receives Products Cost and KPI Assessment Inventory Aging We analyzed one supply chain of Chiquita s Fresh Express line, packaged green salad. We chose this supply chain with Chiquita s input because it represents a large portion of Fresh Express volume and possesses common characteristics shared by Fresh Express other supply chains. The physical structure of this supply chain (one plant, followed by one DC, followed by multiple retailers) is common for Chiquita. Therefore, we expect the results of our research to be applicable to Fresh Express other supply chains and to provide insights to other companies that have similar supply chain structures. Simulation Model After reviewing relevant literature on various approaches to perishable inventory management, we chose to develop a simulation model for our research. Simulation modeling is a cost efficient way to understand the behavior of a complex model, because the user can study the results under Step 5 Step 6 Step 7 Step 8 Figure 1 Overview of Conceptual Flow Input Parameters and Assumptions After building and validating the conceptual model, we transformed the conceptual model into a simulation model using Arena simulation software. In our simulation model, we had the following assumptions: no raw material shortage or capacity constraints at the plant, stochastic and normally distributed demand, First-In-First-Out policy at the DC and retail stores, transportation lead time from the DC to retail stores is always one day, and transportation lead time from the plant to the DC is deterministic and constant for each simulation. In our simulation model, we considered the following input parameters: target days on-hand (DOH) inventory level, shrinkage probability for each day (up to 14 days), forecast error, product unit cost, lost

3 1 Chiquita/DC 2 Calculation in Excel Calculation Within Arena Model Daily Point of Sale At Find Mean & Standard Deviation Daily Product Demand by Consumers at at Add 2.5 day Target Days On Hand, 2.5 Order Up To Level at Order Quantity For DC Aggregate Daily Product Demand at s at DC Add 6 day Order Up To Level, (5+1) Order Up To Level at DC Order Quantity For Plant Daily Point of Sale At Daily Product Demand by Consumers at at Add 2.5 day Target Days On Hand, 2.5 Order Up To Level at Order Quantity For DC Find Mean & Standard Deviation Figure 2 Information Flow in the Model sales cost, inventory holding cost, shrinkage cost, transportation lead time, and production lead time. Figure 2 shows how the input parameters are used in the model and the flow of information in the supply chain. We generated the random daily demands for each retail store using the mean and standard deviation obtained from the Point-Of-Sale data. We then applied the forecast error to this demand and errantly obtained the demand forecast. The demand forecast was used to calculate the order up to level (OUL) for each retail store. For the DC, the daily demand was obtained by aggregating individual orders from each retail store. The forecast error was applied to this aggregated demand from the retail stores to obtain the demand forecast. The order quantity for each echelon was determined by comparing the OUL with its inventory position. Additionally, our review of academic literature on inventory management for perishable products and discussion with Chiquita s supply chain managers showed that none of the analytical methods existed in the literatures could produce the shrinkage profile experienced by Chiquita in reality. We determined that shrinkage probabilities experienced by Chiquita in reality are best fit by an exponential growth model. This means that even for products that are less than 14 days old, an exponentially increasing fraction of inventory is discarded at the end of each day due to perishability. We used Excel Solver to obtain these exponentially growing shrinkage probabilities. We ran the simulation for the base scenario with these shrinkage probabilities and validated that these probabilities indeed produce aggregated total shrinkage volume at the retail stores within the valid range observed by Chiquita. Simulation Results After we verified and validated our simulation model, we conducted tests to determine how different target DOH inventory levels at the distribution center (DC) and retail store would impact total relevant costs and service levels. We tested nine and ten different target on-hand inventory levels in increments of ½ integer values for the DC and retail stores respectively, a total combination of 90 different scenarios. We tested increments of ½ integer values because those are the smallest possible increments that Chiquita operates with. For each scenario, we conducted the simulation for 20 replications, and each replication consisting of 365 days. Tables 1 and 2 present the results for total relevant cost and item fill rates (IFR) for 90 different combinations of target DOH inventory levels at the DC and retail stores. The feasible solutions (that meet the required service level of 95%) are highlighted in green, the results of the current inventory policy are marked in red, and the results of the optimal inventory policy are marked in blue. The results of simulations show that lowering the target on-hand inventory levels at the DC and retail stores to 0.5 and 1.5 days respectively reduces total relevant costs by 31% and maintains high service

4 levels above the required service level of 95% at each echelon. The optimal inventory policy improves the IFR at the DC from 91.74% to 95.24%, while the IFR at the retail stores decreases from 99% to 97%. Although the IFR at the retail stores decreases by approximately 2%, 97% is still considered a high customer service level. The 31% cost savings are significant enough for Chiquita and the retail stores to consider replacing the current inventory policy with the optimal inventory policy obtained through our research. The majority of the cost reduction comes from the reductions in inventory holding costs and inventory shrinkage costs across the system, especially at the retail stores. Table 1 Total Relevant Cost for the System s Target DOH $305 $145 $87 $92 $101 $111 $118 $125 $ $322 $159 $88 $95 $108 $120 $130 $138 $ $342 $178 $102 $111 $126 $140 $153 $165 $ $359 $194 $117 $127 $143 $159 $174 $188 $202 DC 2.0 $380 $214 $134 $144 $161 $179 $195 $211 $ $398 $231 $150 $160 $178 $196 $214 $231 $ $418 $251 $168 $177 $195 $215 $233 $251 $ $436 $269 $184 $194 $212 $232 $251 $270 $ $456 $290 $203 $212 $230 $249 $268 $288 $ $474 $308 $220 $229 $246 $266 $285 $306 $323 Table 2 Item Fill Rate at the DC s Target DOH DC Sensitivity Analysis We tested the sensitivity of the multi-echelon system under the optimal inventory policy with respect to forecast error at the DC, forecast error at the retail stores, and transportation lead time from the plant to the DC. As illustrated in Figures 3 and 4 below, the total relevant cost in the system is more sensitive to the forecast errors at the DC than to the forecast error at the retail stores. The current forecast error at the DC and retail stores is 25%. The sensitivity analysis reveals that increasing the forecast accuracy at the DC by 20% would reduce total relevant costs in the system by 16%. Increasing the forecast accuracy by 20% at the retail stores would reduce the total relevant costs in the system by 7%. Total Relevant Costs Total Relevant Cost $180 $160 $140 $120 $100 $80 $60 $40 $20 $0 $180 $160 $140 $120 $100 $80 $60 $40 $20 $0 Forecast Error vs. Total Relevant Cost for the System Forecast Error Forecast Error vs. Total Relevant Cost for the DC SystemCost/ DCForecastError SystemCost/ RetailForecastError Figure 3 Forecast Error vs. Total Relevant Cost for the System DCCost/ DCForecastError DCCost/ RetailForecastError Forecast Error Figure 4 Forecast Error vs. Total Relevant Cost for the DC Figures 5 and 6 show that the IFR at the DC is more sensitive to the forecast error than to the IFR at the retail stores. The IFRs at both the DC and the retail stores decrease as the DC s forecast accuracy deteriorates beyond 25%. This occurs because the OUL for the DC, which is dependent upon the DC s forecast, becomes less accurate and the shrinkage volume at the DC increases as DC s forecast accuracy deteriorates. This affects the availability of the inventory at the DC and hence its ability to fulfill orders received from the retail stores. For the DC s forecast error less than 25%, the inventory level at the DC is reduced but retail stores still have their own forecast error, so retail stores still tend to order more from the DC. Thus the IFR at the DC deteriorates because the retail stores order more than they need to, suggesting the presence of the bullwhip phenomenon. Item Fill Rate % Forecast Error at the DC vs. Item Fill Rate DC Forecast Error Figure 5 Forecast Error at the DC vs. Item Fill Rate DCIFR Retail1IFR Retail2IFR

5 Overall the retail stores forecast error has a greater impact than the DC s forecast error on the DC s IFR, as depicted in Figure 5 below. Even if the forecast accuracy at the DC is improved, if the retail stores have poor forecast accuracy then the DC s IFR will still suffer due to the reasons mentioned above. Item Fill Rate % Forecast Error at s Vs Item Fill Rate Retail Forecast Error DCIFR Retail1IFR Retail2IFR Figure 6 Forecast Error at the s vs. Item Fill Rate Additional sensitivity analysis with respect to the transportation lead time revealed that by increasing the transportation lead time from the plant to the DC by 1or 2 days would increase the total relevant cost in the system by 11% or 17% respectively. Insights Our simulation model considered forecast accuracy, inventory holding cost, product perishability, lost sales and inventory shrinkage costs and determined that the optimal inventory policy would be to set the target DOH inventory to be 0.5 days at the DC and 1.5 days at the retail stores. Comparing the results for the optimal inventory policy and the current inventory policy, the optimal inventory policy reduces the total relevant cost by 31%, reduces the shrinkage to 4.5%, and maintains the IFR above the required 95% at all echelons. The dynamics of the demand variability, the forecast errors, and the shrinkage in the multi-echelon system amplifies as the target DOH inventory at the retail stores increases, deteriorating the IFR at the DC. Based on our results, we recommend the retail store to keep its inventory less than two days to minimize the upstream impact of the demand variability, the forecast errors, and the shrinkage. impacted by the orders and forecasts created by the retail stores. The IFR at the DC is more sensitive to the forecast error at the retail stores than to the DC s forecast error. To maximize the cost savings and improve the performance of the whole system, we recommend that Chiquita work with the retail stores to simultaneously improve the forecast accuracy at both the DC and the retail stores. The system is sensitive to the transportation lead time since the transportation lead time directly impacts the products lifetime at the DC and at the retail stores. Our model demonstrates that as the transportation lead time increases, the shrinkage in the system increases significantly. This further results in the increased lost sales at the retail stores and reduced IFR at the retail stores and the DC. Since the system is sensitive to the transportation lead time, Chiquita should not relax the current lead time and should continue to keep the transportation lead time as short as possible to maximize the products available lifetime at the DC and the retail stores. Summary We concluded that for multi-echelon perishable inventory management problems, simulation can be extremely helpful. Our research demonstrated that simulation modeling can quantify various trade-offs involved in making inventory management decisions for perishable products. It is extremely valuable to simulate reality and test the sensitivity of the system before decisions are made by managers. Simulation modeling can lead to optimal solutions that would reduce the system costs significantly while improving the system performance significantly. Our research not only answered the three key questions originally posed by Chiquita, but also provided a simulation model that enables Chiquita to analyze different scenarios in order to make effective strategic decisions for its multi-echelon supply chains. We found that the system is very sensitive to forecast errors. In general, as the forecast accuracy deteriorates, the IFR at the DC and the retail stores decreases and the relevant cost of the system increases. Additionally, improving the forecast accuracy at the DC alone may not address the issue sufficiently because the DC s inventory is ultimately