MACHINE LEARNING APPROACH FOR RETAIL DEMAND FORECASTING: INTEGRATING FEATURE ENGINEERING WITH STACKED TREE-BASED ENSEMBLE AND DEEP LEARNING MODEL FOR IMPROVED ACCURACY

Abstract

Retail demand forecasting plays a critical role in supply chain management by enabling businesses to predict future sales, manage inventory efficiently, and enhance production planning. With the advancement of machine learning, particularly tree-based ensemble methods and deep learning techniques, traditional forecasting systems have evolved to better handle the complex and non linear patterns present in retail data. This study evaluates the forecasting performance of a stacked ensemble comprising tree-based models—Random Forest, XGBoost, LightGBM, and CatBoost— using Gradient Boost as the meta-learner, in comparison to an artificial neural network (ANN), a widely used deep learning model. The analysis is conducted on a five-year dataset covering multiple stores and products, using comprehensive feature engineering methods such as lag variables, rolling windows, month-over-month sales growth, and interaction terms to uncover significant temporal and cross-sectional patterns. Forecasts are generated for a three-month horizon to aid inventory control and production planning. An ANOVA test indicated that approximately 71% of the sales variance could be explained by engineered features, validating its effectiveness. The stacked ensemble model significantly outperformed the ANN, achieving a maximum R² value of 0.994 compared to 0.924 from the ANN. Moreover, the ensemble approach surpassed the performance of individual models, with the best-performing tree-based model incorporated into the stack. Overall, the study highlights that when supported by effective feature engineering, tree-based stacking ensembles offer superior accuracy in capturing non-linear relationships in retail demand forecasting, and statistical methods can be used to make decision for feature engineering to improve the forecast.