Domain Adaptation for Retail Demand Prediction

Abstract

Predicting the demand of products in the retail industry is a complex task, especially when there are changes in the market. This paper examines three such market shifts in the retail industry: the COVID-19 pandemic, opening a new store, and introducing a new product. Our study found that the accuracy of demand prediction models decreases after these market shifts. To address this problem, we propose the use of domain adaptation methods, such as Frustratingly Easy and Kernel Mean Matching, to improve the accuracy of predictions by utilizing data from before the changes and adapting to the data after the changes. We show that using a pairing technique can further enhance prediction accuracy. Two retail demand forecasting models, XGBoost and Transformers, were assessed and XGBoost was found to be more effective. We demonstrate the effectiveness of the domain adaptation methods on a real-world case using point-of-sale data from 89 locations of Alimentation Couche-Tard convenience stores in Montreal between 2019-07 and 2021-02 focusing on the two best-selling product categories of coffee and energy drinks.