A Hybrid Approach for Processing External Data and Forecasting Sales Demand Using LSTM: A Case Study of Rossman

This study presents a novel hybrid approach for forecasting sales demand by effectively integrating external data —specifically weather conditions —into a deep learning framework. In the first stage, weather data (including temperature and categorical weather events) is processed using both one-hot encoding and a probability-based scoring mechanism. A Random Forest regressor is employed to quantify the impact of various weather events on sales, enabling the transformation of qualitative weather conditions into meaningful quantitative features. Temperature values are evaluated using a normal distribution-based scoring function, and the resultant scores are combined with the adjusted weather importance to form a unified external feature. In the subsequent stage, this composite feature, along with internal sales and state-level information, is utilized as input to a Long Short-Term Memory (LSTM) network for time series forecasting. When applied to the Rossmann case study, the proposed method demonstrated an approximate ۵% improvement in forecasting accuracy compared to conventional LSTM models that rely on standard feature transformation techniques. These findings underscore the potential of the hybrid approach to capture complex external influences, thereby enhancing predictive performance in retail demand forecasting.