Hybrid Models for Retail Demand Forecasting: Integrating Classical Time-series and Machine Learning Approaches

This paper presents an idea and implementation of using hybrid models in forecasting retail demand by integrating conventional approaches with machine learning. The objectives are to develop methodologies for building composite structures, sequential residual models, group stacking/blending, and orderly mixtures, exemplified by combinations like ARIMA–LSTM and Prophet–LightGBM, and to explore their primary pros and cons in meeting the changing demands of retail forecasting. The present work gains critical importance against a backdrop of markedly increased unpredictability in human behaviour—manifested by rapid sales fluctuations and heightened pressures throughout global supply chains. In such a context, enhancements in forecasting accuracy become directly linked to both revenue optimisation and the reduction of safety-stock expenditure. A critical appraisal of over twenty scholarly sources underpins the novelty of this research, which proposes a comprehensive methodological framework for organising the MLOps lifecycle in hybrid system architectures.The study justifies the separation of linear and nonlinear forecast components, which enables a reduction of MAPE by double-digit percentages and a decrease in error variance for cold SKUs while preserving the interpretability of the statistical component. The main conclusions are as follows: hybrid models demonstrate double-digit reductions in MAPE, RMSE, and WRMSSE compared to standalone algorithms. For practitioners, the resulting modular architecture simplifies maintenance and offers a scalable solution, while integrated CI/CD and CT loops ensure reliability and rapid response to data drift. The modularity of the architecture simplifies maintenance and scaling. Additionally, CI/CD and CT loops ensure reliability and rapid response to data drift. The key factor for success is the balance between explainability, response time, and the incorporation of spatial correlations. This paper will be useful for researchers and practitioners in Data Science, MLOps engineers, analysts, and operational managers of retail chains.