A Machine Learning Framework: E-Commerce Retail Sales Forecasting with External Weather Features
TL;DR
Historical sales information was the primary predictive signal, while weather provided supplementary information, and an interpretable forecasting framework for daily-location sales revenue can support revenue-oriented forecasting and help assess the additional value of incorporating weather information.
Abstract
Background: Accurate sales forecasting supports inventory planning, resource allocation, revenue management, and operational decision-making. However, the incremental predictive value of external factors such as weather, beyond historical sales information, remains important to assess. This study evaluates an interpretable forecasting framework for daily-location sales revenue and examines the relative contributions of historical and weather information. Methods: Daily sales data were integrated with weather observations and temporal lag features. Random Forest, Extreme Gradient Boosting, CatBoost, and Stacked Long Short-Term Memory (LSTM) models were used for forecasting. Performance was assessed using mean absolute error, root mean squared error, symmetric mean absolute percentage error, and R2. Results: Stacked LSTM achieved the strongest overall test performance (MAE = 28.50, RMSE = 57.04, sMAPE = 5.82%, R2 = 0.972). Feature-importance analysis indicated that historical sales, particularly one-day sales lag, provided a stronger predictive signal than individual weather variables. Conclusions: Historical sales information was the primary predictive signal, while weather provided supplementary information. From a practical perspective, the framework can support revenue-oriented forecasting and help assess the additional value of incorporating weather information. This study also contributes to forecasting research by distinguishing the predictive role of historical sales patterns from the incremental contribution of external contextual variables.