Comparative Analysis of Demand Forecasting Models: A Time Series Case Study

(1) Background: Accurate demand forecasting is essential for supply chain management, guiding operational planning, decision-making, and resource allocation. Evaluating classical statistical and AI approaches is crucial for selecting suitable forecasting techniques. (2) Methods: This case study assessed five approaches—classical statistical algorithms, machine learning models, and a hybrid method—using consumption data from a retail chain. Performance was measured by predictive accuracy (sMAPE), training and testing latency, memory usage, and storage requirements. (3) Results: XGBoost achieved the best balance, with low predictive error (sMAPE = 1.03%), minimal latency (79.3 ms), and moderate memory usage (359.58 MiB). SARIMAX and Holt-Winters remained competitive, providing robust, interpretable, and efficient predictions. LSTM showed higher computational demands (867.52 MiB) and lower accuracy (sMAPE = 3.7), while the Holt-Winters plus XGBoost hybrid did not outperform individual methods (sMAPE = 1.36). Classical algorithms produced optimistic forecasts, supporting make-to-stock strategies, whereas AI models were more conservative, aligned with a following-demand approach. (4) Conclusions: Model selection should consider predictive accuracy, operational costs, product characteristics, and organizational strategies. Results are dataset-specific and cannot be generalized. Limitations include the absence of exogenous variables. Traditional statistical methods remain competitive, interpretable, and efficient against AI approaches in structured demand forecasting.