invent analytics size and prepack optimization white paper

INVENT ANALYTICS SIZE AND PREPACK
OPTIMIZATION WHITE PAPER
Gürhan Kök, Invent Analytics, Duke University Özgür Emre Sivrikaya, Invent Analytics, Bogazici University Yunus Emre Koç, DeFacto Engin Şahin, DeFacto, 2014 ABSTRACT Pre-­‐pack optimization is a common problem in fashion retailing. Retailers pack multiple sizes of the same product (style) into the same package to minimize warehouse costs. We develop a stochastic inventory model based evaluation system for determining the optimal package configurations and the procurement amount of each package type. Live controlled experiments demonstrate a 10% increase in gross margin after taking into account a 5% increase in sales and slight increase in logistics costs. Size and Prepack Optimization to Minimize Lost Sales and Logistics Costs at a Fashion Retailer SUMMARY This paper describes a supply chain improvement Project at DeFacto (Ozon Giyim), one of the leading fashion retailers in Turkey with 220 stores and $500 million in annual revenues. Pre-­‐pack optimization is a common problem in fashion retailing. Retailers pack multiple sizes of the same product (style) or even multiple products into the same package to minimize warehouse picking and shipping costs. However, due to 1) different size distributions of different stores and 2) high uncertainty in demand, working with one single package leads to significant inflexibility in the distribution system and may result in high lost sales for some store-­‐product combinations and high excess inventory levels for some others. In this project, Ozon Giyim and Invent Analytics, an academic consulting firm, jointly developed a stochastic inventory model based evaluation system for determining the optimal package configurations, the optimal procurement amount of each package type and and an initial allocation of packages to the retail stores. Live controlled experiments demonstrated the impact of the recommended solutions: a 10% increase in gross margin after taking into account a 5% increase in overall sales and slight increase in logistic operational costs. COMPANY BACKGROUND Ozon Giyim, founded in 2003, entered fast fashion retail market in 2005 after opening two new stores introducing its new brand, DeFacto. DeFacto grew its business with 20 stores in 2007, 90 stores in 2009 and 220 stores in 2014. Until 2012, DeFacto had been focused on domestic market. Then starting from 2012, DeFacto entered new markets including Kazakhstan, Iraq, Iran and Egypt with 30 new stores until the end of 2014. DaFacto’s fast growing business leads to new challenges in its supply chain operations. PROJECT DESCRIPTION Early on, store locations were close to each other, therefore product sizes were following similar distributions and demand were met with one type of package for every product. Initial allocation and replenishment of all products were handled with this package type. However, with increasing number of new locations, this policy ended with stock imbalances between the stores. In some locations small sizes (XS, S, M) were demanded most, whereas in other locations larger sizes were sold out. Therefore, transfer costs between the stores were increased to balance this demand fluctuations. Demand distribution of stores can be found in the figure below. Hence, DeFacto decided to change its one-­‐type package policy to decrease its storage, operation and transportation costs. N1: 39
N2: 50
N3: 77
SOLUTION 1. Forecasting: Pre-­‐pack optimization problem offers different package types to distribute optimal number of products to each store. Inputs of he problem are sales forecasts and logistics costs. Logistic costs are deterministic and it is not possible to change these costs. However, forecast quality affects the success of the solution substantially. Thus, at first, a product-­‐store-­‐season based forecast model was developed. Historical data regarding sales, stock, price and important days like national and religious holidays were considered as input in the linear regression based forecast methodology. To measure the forecast quality, standard methods like mean absolute deviation (MAD) and mean squared error (MSE) were used. 2. Mathematical Programming Model: After the validation of the forecast model, a mathematical model was developed to determine the optimal package configurations to minimize the total mismatch between stores’ demand and the planned shipments to the store. In this process DeFacto’s certain constraints had to be considered. These constraints were as follows: At most fifteen products can be handled with one package. Every package has to include only consecutive sizes. For example S and L cannot be carried with one package unless it contains a M sized product. It is allowed to create a package type which contains only one size of each product. However, a package with only extreme sizes (i.e. package including only XS) of certain products were not allowed. At most three products with same size can be carried with one package. This mathematical problem is an NP-­‐hard problem. However, it is possible to solve the problem in a reasonable amount of time, since the problem is separable for each product. The model was integrated into DeFacto’s system and solved using CPLEX. 3. Simulation: To validate the mathematical model’s results, a simulation of the previous season was developed using C#. At first, using forecast model, sales forecasts were generated. Then mathematical model was run using these forecsats and package types were determined. Finally sales distribution of stores was determined and simulated sales were generated randomly for each store. Then these packages were distributed to stores using DeFacto’s allocation algorithm. According to simulation results, gross margin was increased by 5% -­‐ 10%. 4. Live Controlled Experiments: In the current season (Spring -­‐ Summer 2014), a real-­‐time test of the pre-­‐pack optimization project is being carried out. Stores are divided into two groups by a matching algorithm. Stores were classified based on the following variables: Climate region, income level of client base, Street vs mall. The stores in each cluster were randomly assigned to either the test or the control group. There are 76 stores in each group for Product 1 and 74 stores in each group for Product 2 and the total revenue and sell through of the test and control groups are within 1% of each other. Test group is managed by new pre-­‐pack optimization method, whereas the control sample is managed with the previous one-­‐type package policy. Demands of the two groups can be found in the figure below. According to the results, the test sample generated 10% higher gross margin. Results of the different packaging policies can be found below. Product
Store Group
Store Count
Total Demand
Sales
Cost of Stock Out
Cost of Stocj Out %
Revenue (witout VAT)
Cost of Goods Sold
Total Logistics Cost
Gross Margin
Product 1
Referance
Test
76
76
2.087
2.042
1.598
1.666
489
376
23%
18%
103.559
107.966
43.945
45.815
813
895
58.801
61.256
Product 2
Referance
Test
74
74
1.453
1.450
1.202
1.276
251
174
17%
12%
77.896
82.692
33.055
35.090
652
673
44.189
46.929
RESULT DeFacto has started implementing the algorithm across all categories and buyer groups. The full scale implementation has been done internally enabling buyers and category managers to see the package configurations and the purchase quantities of each package type and the simulation results of the recommended solution.