Optimizing Retail Operations through Hybrid Machine Learning and Deep Learning Techniques in Demand Forecasting

This paper introduces StackRetailNet, an innovative meta-learning-based hybrid model designed for retail demand forecasting and inventory classification. Compared to simple fusion methods like weighted averaging, StackRetailNet uses a stacked ensemble method that combines various machine learning algorithms (Light Gradient Boosting Machine, Random Forest, Extreme Gradient Boosting) and deep learning models (Long Short-Term Memory, Recurrent Neural Network, Transformer, Feed Forward Neural Network). The meta-learner adaptively learns to optimize predictive performance out of the base model outputs. In the proposed pipeline, the regression stage first predicts sales volumes, and these forecasted values are then provided as inputs to a deep neural network-based classification model for forecast-aware inventory planning. This sequential structure ensures that inventory categorization incorporates projected future demand rather than relying solely on historical data. Experimental evaluation of StackRetailNet yielded an accuracy of 98.95%, precision of 96.25%, and recall of 95.12%, thereby demonstrating significant improvements over traditional methods. The advanced feature engineering (optimized rolling windows, external covariates, and feature importance-based selection) makes the model more robust. Therefore, this study offers key insights for future research to improve forecasting model selection, explanation, data quality, and computational solutions for retail demand forecasting accuracy.