Multi-Source Data Fusion for Short-Term Demand Forecasting of Seasonal Retail Products: An Empirical Study Using Weather and Social Media Signals

Recent advances in machine learning have enabled retailers to leverage diverse data streams for demand prediction, yet most systems continue to rely primarily on historical sales. In this study, we address the challenge of seasonal demand forecasting by integrating meteorological observations, social media activity, and economic indicators with point-of-sale transactions. We propose a hierarchical ensemble that combines gradient boosting machines with bidirectional LSTMs, where component weights adapt according to product category and forecast horizon. Through mutual information screening, we reduce feature dimensionality from 123 to category-specific subsets averaging 38 features, lowering computational costs by 60% while maintaining predictive accuracy. Our approach requires only standard training of the base models along with a lightweight meta-learner, without necessitating end-to-end joint training. Evaluation on 10.3 million transactions from three U.S. retail chains demonstrates that our method achieves a mean absolute percentage error of 19.6%, compared to 37.5% for seasonal naive baselines. Multi-source fusion thus provides a practical pathway to more accurate retail forecasting, without the complexity associated with end-to-end deep learning systems. Crucially, this approach relies on lightweight feature learning and selection techniques combined with meta-learner weighting, rather than on end-to-end joint representation training.