Cross-channel predictive analytics for retail distribution decisions

Abstract

Distribution demand forecasting at Zara currently considers historical sales of products modified by expert knowledge inputs in an algorithm developed to calculate the shipment required to meet demand for the next sales period. In 2010, the introduction of Zara.com provided customers an additional channel to complete purchases and interact with the brand while providing Zara significant insight into changing customer preferences to supplement the expert knowledge of the Zara team. This thesis investigates the utility of the data collected in the online sales channel for increasing the accuracy of the distribution demand forecasts. Two forecast types are considered: Initial Shipments for which no historical data exists, and Replenishment Shipments which have historical data. Forecasts are performed for both brick-and-mortar and e-commerce sales channels to demonstrate cross-channel utility of the data. The study presents a review of available datasets to identify those of potential interest and describes meaningful features engineered from raw datasets. By applying machine learning algorithms, significant features are identified and a predictive model is developed demonstrating significant WMAPE improvement for initial shipments to brick-and-mortar stores ( 0.23), moderate improvement for replenishment shipments to e-commerce ( 0.05) and limited improvement for replenishments to brick-and-mortar stores (<0.04). The results of this study demonstrate the potential for significant reduction of inventory requirements to maintain customer service levels and provides a baseline for future cross-channel forecasting work.