Explainable time-series forecasting with sampling-free SHAP for Transformers

Speed enables practical deployment: SHAPformer generates explanations in under one second—several orders of magnitude faster than existing SHAP methods—making explainability feasible for real-time decision-making systems rather than just post-hoc analysis.

Better handling of real-world data relationships: Traditional SHAP assumes features are independent when creating counterfactual scenarios, which is unrealistic for time series. SHAPformer's sampling-free approach respects actual feature dependencies, potentially providing more accurate explanations of what the model actually learned.

Explanations validated against ground truth: On synthetic data with known correct explanations, SHAPformer's outputs matched the ground truth, suggesting the method reliably identifies which factors genuinely drive predictions rather than producing spurious explanations.

Actionable insights for domain experts: When applied to electrical load forecasting, the model revealed interpretable patterns (e.g., past load as the key predictor, distinct behavior during Christmas), demonstrating that faster explanations can surface meaningful business insights that support human decision-making.

Maintains competitive accuracy while adding transparency: SHAPformer achieves predictive performance comparable to standard Transformers while providing explanations, addressing a key tension in responsible AI: users don't have to sacrifice model performance to understand how it works.