Time series forecasting models are increasingly deployed in continuously evolving environments. After a forecast is issued, subsequent observations gradually arrive and expose the model's failures. Existing approaches mainly treat these observations as supervision for parameter updates, without clarifying whether verified forecasting errors can serve as causal feedback to correct future predictions. To address this gap we introduce Error Feedback Calibration (EFC), a fully test‑time deterministic post‑hoc correction framework.
EFC first extracts revealed errors from overlapping historical forecasts to build a Causal Error Diagonal (CED), offering a fresh view of the frozen forecaster's recent failure behavior. Then Cumulative Error Transport (CET) maps this error state to future residuals via cumulative estimation, while out‑of‑sample temporal validation controls the correction strength. Temporal Residual Learning (TRL) further predicts only the residual structure left after CET, avoiding a full re‑learning of the forecasting error.
Extensive experiments across various forecasting backbones and datasets demonstrate that EFC broadly improves forecasting accuracy and substantially mitigates harmful over‑correction.
Review