A hybrid learning framework for forecasting uncertainty and adaptive inventory planning in retail supply chains

Demand forecasting and quantification of uncertainty is an essential asset of the retail supply chain optimization and risk-based inventory decisions. This study will introduce a new hybrid conditional variance model (combining gradient boosting machines (XGBoost, LightGBM), recurrent neural networks (LSTM-GRU hybrid), and econometric volatility modeling (GARCH) using a stacked ensemble meta-learning method to make retail demand forecasts over multiple horizons. The framework handles important deficiencies of current methods by providing simultaneously high-precision point predictions and probability prediction intervals by conditional estimation of variance. The M5 Walmart benchmark dataset of 8000 high-volume product time series including all features engineered in terms of 58 time, statistic, price and event dimensions are empirically validated. Stacked ensemble architecture has high predictive work at R 2 = 0.9681, root mean squared = 1.48 units and mean absolute error = 0.77 units, which is significantly better than base models. Integrated GARCH(1,1) component effectively explains forecast residual volatility whose mean conditional variance is 2.82 square units, which allows it to construct dynamically adaptive 95% confidence intervals. Forecast shift analysis shows average magnitude of day-to-day revision of 3.21 units with great correlation between the magnitude of the predicted variance and the actual forecast volatility. The proposed framework offers supply chain practitioners actionable probabilistic predictions to aid risk-conscious inventory location and adaptive safety inventory determination, which is a major improvement over traditional point estimation techniques. • Develop an analytical framework integrating learning and econometric models for retail forecasting. • Achieve superior predictive accuracy across thousands of retail product time series. • Quantify forecast uncertainty using conditional variance modeling for adaptive decision making. • Analyze forecast shifts to connect prediction volatility with daily demand variation. • Enable risk-aware inventory and replenishment decisions through probabilistic analytics.