SALES FORECASTING AND DEMAND PREDICTION THROUGH TIME SERIES ANALYSIS AND MACHINE LEARNING

Abstract

Accurate sales forecasting is central to effective inventory planning, resource allocation, and supply chain management. Classical time series models, such as ARIMA and Holt-Winters, are widely used due to their interpretability and strong performance on stationary or seasonal data. However, these models often fall short in capturing nonlinear dynamics and abrupt changes in real-world sales patterns, which are influenced by promotional events and holidays. In this study, we propose a hybrid forecasting framework that integrates statistical time series decomposition with machine learning techniques, specifically Long Short-Term Memory (LSTM) networks and Gradient Boosting Regression. Using five years of daily retail sales data enriched with external variables, we compare ARIMA, Holt–Winters, GBR, LSTM, and a hybrid ARIMA–LSTM method. We evaluate forecasts using RMSE, MAPE, and R² under a rolling origin validation scheme. Results show that the hybrid model reduces MAPE by up to 18 % relative to classical methods, achieving a balance between interpretability and predictive performance. Our findings underscore the value of combining classical and machine-learning models for robust demand prediction in retail.