When Do Corrective Features Help? An Agent for Corrective Feature Discovery on Black-Box Forecasters

The authors argue that frozen forecasters often exhibit recurring, structured failures, and fine-tuning to fix them is costly. CRAFTER mines interpretable features of the forecaster's residual and uses them to drive a lightweight corrector, avoiding full retraining. The paper investigates when corrective features are likely to be helpful, providing theoretical and empirical guidance. This approach is particularly relevant for black-box or expensive models where access to weights is limited. The work is available on arXiv (2608.05207) and presents a new agent-based method for post-hoc model improvement.