Combined method for forecasting food product demand based on an ensemble of LSTM network and SARIMA model for integration into ERP systems of Russian retail

The purpose of the study. To improve the accuracy of food demand forecasting under conditions of high volatility and complex seasonality for integration into inventory management modules of ERP systems, including Russian solutions based on the 1C platform. Materials and methods. A hybrid method is proposed based on an ensemble of SARIMA (Seasonal Autoregressive Integrated Moving Ave¬rage) and a multilayer LSTM network (Long Short-Term Memory). Model weights are determined adaptively based on validation error. The experiment was conducted on real data from the M5 Forecasting Competition (Walmart), covering demand time series for 120 food products. Evaluation metrics included MAE, RMSE, MAPE, and the Diebold–Mariano test. Results. The proposed ensemble reduces Mean Absolute Percentage Error (MAPE) to 52.96 % – 1.1 % better than SARIMA and 14.0 % better than LSTM. Statistical significance of the improvement was confirmed by the Diebold–Mariano test (p < 0.001). The combination of SARIMA’s interpretability and LSTM’s nonlinear flexibility provides robustness to outliers and higher accuracy during sharp demand fluctuations (e.g., before holidays). The practical value of the study lies in the possibility of reducing the level of shortages and excess stocks through a more accurate demand forecast. Conclusion. The developed method shows strong potential for integration into ERP systems used in Russian retail, where a balance between accuracy, interpretability, and automation is essential. The results support the practical adoption of the ensemble in automated procurement and inventory planning modules.