Joint handling of normal, input-only, and input-output anomalies at test time

Establish a principled time-series forecasting approach that, at test time, jointly handles (i) normal conditions, (ii) input-only anomalies confined to the historical input window, and (iii) input-output anomalies that begin in the input window and persist into the prediction horizon, ensuring reliable forecasting under these scenarios.

Background

The paper distinguishes three test-time scenarios in multivariate time-series forecasting: normal conditions, input-only anomalies that should not affect the forecast, and input-output anomalies whose effects persist into the prediction horizon. Existing models often treat all deviations uniformly, either overreacting to noise or failing to adapt to persistent regime shifts.

Prior robust forecasting work largely focuses on anomalies in training data or pointwise corruptions, leaving the challenge of handling unseen anomalies at test time without a unified framework. The authors emphasize that addressing this joint setting is crucial for real-world deployment, where transient disturbances and persistent shifts occur unpredictably.