Adaptive LSTM Model for Predicting Time Series of Sales in the Presence of Conceptual Data Drift

The paper examines sales forecasting when the relationship between demand and its drivers changes after model deployment. An adaptive framework is proposed that combines a two-layer LSTM network, ADWIN-based monitoring of forecast errors, and two parameter-update modes. The experiment uses a four-year daily sales dataset for five agri-food product groups and separately represents covariate shift, abrupt concept drift, gradual drift, and a recurring regime. Across six independent experimental runs, the adaptive LSTM achieved a mean MAE of 12.27, improving the static LSTM by 4.17%. It was the most accurate model during gradual drift, while scheduled retraining retained a small advantage over the full horizon. The findings suggest that drift-aware updating should complement rather than mechanically replace regular model maintenance.