An Adaptive Hybrid ARIMA-XGBoost Framework with Real-Time Streaming for Retail Demand Forecasting

Forecasting retail demand is a difficult undertaking since customer demand is influenced by temporal dependencies, changing sales patterns, nonlinear relationships, and differences between products and stores. This study introduces an adaptive hybrid forecasting approach that integrates Autoregressive Integrated Moving Average (ARIMA) with Extreme Gradient Boosting (XGBoost) in order to predict retail demand. ARIMA is employed to model the temporal structure of each individual demand series, and XGBoost is used to learn the residual patterns that are left after the statistical forecast has been carried out. An adaptive rolling forecasting procedure is included so that recent observations can be utilized to update the later predictions. The framework is assessed using the M5 retail forecasting dataset, keeping the demand data at the individual store-product level rather than aggregating the series. In order to apply the forecasting process in a real-time environment, Apache Kafka is used for continuous data ingestion and processing, and a Streamlit dashboard is provided to visualize the predictions, the errors, and the forecasting performance. The experimental results indicate an RMSE of 1.9796, a MAE of 0.9799, and a WAPE of 72.89%, which shows an 8.73% reduction in RMSE compared to the ARIMA baseline forecast. The results prove that combining statistical forecasting with machine learning-based residual correction and streaming infrastructure offers a practical solution for responsive retail demand forecasting.