Domain Adaptation For Retail Demand Prediction

"Demand Forecasting is an important tool in many industries including retail. Althoughmany approaches have been developed to accurately predict the demand of productsbased on their historical sales data, demand prediction is still a complex issue especiallywhen there is a domain shift between training and testing data.In this work, we study three examples of domain shifts in the context of retail: outbreak ofthe COVID-19 pandemic, opening a new store, and introducing a new product. We firstshow that the accuracy of demand prediction models suffers after each sudden change.Then, we use domain adaptation methods, such as Frustratingly Easy (FE) and KernelMean Matching (KMM) to help improve the demand prediction accuracy by leveragingthe available data from the period before the shift (source domain) and adapting it to thedata after the shift (target domain). Additionally, we show that using a pairing techniquefurther helps improve the prediction accuracy.We use two methods as our base forecasting model: XGBoost and Transformers, and weshow that in the context of our data, it is better to use XGBoost.Our dataset comprises of point-of-sales data from 89 locations of Alimentation Couche-Tard convenient stores in the island of Montreal gathered between 2019-07 and 2021-02.We use product price information in addition to sales information to predict the demandof products in each store. In this study, we focus our attention on the two high-sellingcategories of coffee and energy drinks"--