Application of deep neural networks for sales forecasting: an integrated approach

Abstract

This study investigates the application of deep learning ensembles, combining LSTM, CNN, and MLP architectures, to improve the accuracy of sales forecasting in a technology company. Using a comprehensive dataset covering 2021 to 2024, with more than 10,000 monthly records, a robust temporal validation protocol was implemented. The results indicate that the ensemble model consistently outperforms classical and individual models, showing a sMAPE of 12.3%, MdAPE of 10.5%, WAPE of 15.2%, and an RMSE of 0.28 for the horizon H=1, figures that reflect a better ability to capture complex patterns and temporal dependencies in sales data. These values show a significant improvement over base models such as the naive seasonal model, which exhibits an sMAPE of 25.0% and RMSE of 0.50 for the same horizon. This integrated method offers an effective tool for strategic decision-making and efficient inventory management, although it requires more computational resources and greater difficulty in calibrating and training the models. However, the advantages in accuracy and robustness make the investment worthwhile, positioning deep learning models as a whole as an advanced solution in commercial forecasting according to the related literature in the area.