Explainable Time Series Forecasting

Explainable time series forecasting studies how to show which past observations, variables, contexts, or learned dynamics shape predictions across future horizons. Approaches include post-hoc methods such as SHAP and counterfactuals, interpretable decompositions and symbolic equations, causal or concept-based models, and language-model reasoning. Methods are tested on settings such as energy demand, health, traffic, and industrial forecasting, where temporal dependence, multiple variables, uncertainty, and changing regimes matter. A central challenge is ensuring explanations are faithful and stable—not merely plausible—while preserving forecast accuracy, especially when different horizons rely on different evidence.