In time‑series forecasting, feature attribution must answer two questions that a scalar per‑feature score cannot: when a feature acts within the input window and how its influence varies across the forecast horizon. Existing Shapley explainers treat each feature as a single player, collapsing both axes. We introduce TempShap, a retrospective loss‑based Shapley framework whose players are the individual input cells $(t,i)$. The resulting attribution tensor yields both temporal‑localization and horizon‑impact profiles. Validation is performed on three fronts: synthetic data with planted temporal structures, a tractable exact cell game, and a closed‑form attribution oracle for a linear forecaster. Permutation sharing estimates the cell‑level game under a localized perturbation cost. Under redundant inputs, one‑at‑a‑time scores under‑credit substitutable information, whereas the signed Shapley decomposition preserves total value and provides more reliable held‑out deletion rankings under mean masking. A crafted counterexample isolates the redundancy mechanism. Experiments show a deletion‑ranking advantage across Shapley estimators, forecasting benchmarks, and model families. Hence, TempShap makes fine‑grained Shapley attribution both practical and operationally useful.
Review: TempShap’s cell‑wise player design enables precise timing and horizon‑wise quantification of feature influence, offering a more nuanced and dependable explanation method for time‑series models.