Optimizing Demand Forecasting: Classical Statistical Models vs. AI-Driven Approaches

International competition between retailers is very intense in actual world and the capacity to remain profitable is essential influenced by the ability to forecast the demand as accurately as possible. Also, is very important to minimize stock outages, to reduce the volume of each product in stock, to reduce costs and to increase the revenues. The current work proposes a new approach to forecasting and optimization in the retail environment and introduces an improved method for identifying the optimal algorithm that can help a retailer to get as close as possible to mentioned conditions. The research methodology identifies the optimal forecasting model and then optimize the model results and parameters such as the optimal order quantity, the reordering point, the maximum stock and the safety stock. To obtain the optimal solutions we develop distinct forecasting and optimization algorithms, starting with classical econometric methods like ARIMA, or dynamic optimization, and progressing to more complex approaches, including transformers, neural networks, Fourier transformation, or combined algorithms like Fourier transformation with Prophet. An important challenge was identifying an optimal method for selecting the best operating algorithm, as classical methods like RMSE, MAPE, or MAE failed to identify the most accurate prediction algorithm. In order to test the algorithm’s quality, we test twenty-six distinct forecast algorithms and compared at each product level, identifying Fourier transformation, either on its own or in combination with other algorithms, as a highly effective forecasting model. Finally, the Continuous Ranked Probability Score is proposed as the best model selection strategy.