Simplifying tree-based methods for retail sales forecasting with explanatory variables

Despite being consistently outperformed by machine learning (ML) in forecasting competitions, simple statistical forecasting techniques remain standard in retail. This is partly because, for all their advantages, these top-performing ML methods are often too complex to implement. We have experimented with various tree-based ML methods and find that a ‘simple’ implementation of these can (substantially) outperform traditional forecasting methods while being computationally efficient. Our approach is validated with a dataset of 4,523 products of a leading Belgian retailer containing various explanatory variables (e.g., promotions and national events). Using Shapley values and slightly adjusted tree-based methods, we show that superior performance depends on the availability of explanatory variables and additional feature engineering. For robustness, we show that our findings also hold when using the M5 competition dataset. Extensive numerical experimentation finally shows how the forecast superiority of our proposed framework translates to higher service levels, lower inventory costs, and improvements in the bullwhip of orders and inventory. Our framework, with its excellent performance and scalability to practical forecasting settings, we contribute to the growing body of research aimed at facilitating the higher adoption rate of ML among ‘traditional’ retailers.