Enhancing Sales Forecasting Accuracy in the Presence of Missing Data

Accurate sales forecasting serves as a critical foundation for informed business decision-making and numerous methodologies have been advanced by scholars predominantly rely on complete datasets, however, in practice, uncontrollable variables often result in the presence of missing data, diminishing the applicability of established techniques, so the sales forecasting research based on missing data is of great significance. Some researchers have resorted to employing deletion or interpolation techniques to pre-process incomplete data before conducting sales forecasts, but such completion methods tend to introduce distributional biases, subsequently leading to forecast inaccuracies. In light of these challenges, this study introduces a novel computational framework combining Convolutional Neural Network (CNN), Self-Attention, and Sequence-to-Sequence (Seq2Seq) models, where global features are first extracted via the synergistic application of CNN and Self-Attention mechanisms, then the extracted features are seamlessly integrated with the original dataset, supplanting conventional interpolation strategies, and finally the amalgamated data are inputted into the Seq2Seq model for sales prediction. The proposed methodology is empirically evaluated using sales data from a real-world company, demonstrating commendable predictive performance and underscoring the efficacy of the proposed approach in enabling accurate sales forecasts within the context of missing data scenarios.