AI Agent Improves Black-Box Forecaster Accuracy with Corrective Features.

Pretrained forecasting models, even when frozen, often exhibit systematic and recurring errors that are costly to fix through traditional fine-tuning. This research introduces the concept of corrective feature discovery, which involves mining interpretable features from a frozen forecaster's residual errors to power a lightweight, post-hoc corrector. Unlike conventional automated feature engineering that models the data-generating process, corrective features specifically target the model-failure process. The proposed system, CRAFTER (Corrective Residual Agent with Feature-based Temporal Exploration and Reasoning), keeps the core forecasting backbone frozen. It employs two complementary generators to mine its residual: a compositional search over raw input channels and a large language model (LLM) that suggests named feature combinations, binary flags, and executable code snippets. A validation-grounded gate then accepts or rejects each candidate feature, regardless of its origin. A validation-selected corrector applies the accepted features or leaves the forecast unchanged. This source-agnostic pipeline allows for fair comparison with other feature-engineering systems. Across six public datasets and six different frozen backbones, CRAFTER consistently outperformed all dedicated feature-engineering systems at every feature budget. It roughly doubled the improvement achieved by the corrector alone and reduced the error of the weakest backbones by up to 27%. These gains proved robust across various LLM backends and even when applied on top of already fine-tuned backbones, demonstrating its effectiveness in enhancing forecasting accuracy without complex retraining.