When Do Corrective Features Help? An Agent for Corrective Feature Discovery on Black-Box Forecasters
For practitioners using frozen pretrained forecasters, CRAFTER offers a method to improve performance without fine-tuning, but the gains are domain-specific and incremental over existing feature engineering.
The paper introduces CRAFTER, an agent that discovers corrective features from the residuals of frozen black-box forecasters to improve their performance via a lightweight corrector. Across six datasets and six backbones, CRAFTER outperforms dedicated feature-engineering systems at every feature budget, roughly doubling the improvement of the corrector alone and reducing error of the weakest backbones by up to 27%.
Frozen pretrained forecasters often fail in structured, recurring ways that are costly to repair through fine-tuning. We study corrective feature discovery: mining interpretable features of a frozen forecaster's residual to drive a lightweight post-hoc corrector. Prior automated feature engineering models the data-generating process; corrective features instead model the model-failure process. We present CRAFTER (Corrective Residual Agent with Feature-based Temporal Exploration and Reasoning), which keeps the backbone frozen and mines its residual with two complementary generators: a compositional search over the raw input channels, and a large language model (LLM) that proposes named feature combinations, binary flags, and short executable code. A single validation-grounded gate accepts or rejects every candidate regardless of its origin, and a validation-selected corrector applies the accepted features or leaves the forecast unchanged. This source-agnostic pipeline also allows prior feature-engineering systems to be evaluated under identical conditions, making CRAFTER an instrument for attributing forecast improvements to the feature source alone. Across six public datasets and six frozen backbones, CRAFTER surpasses every dedicated feature-engineering system at every feature budget, roughly doubling the improvement achieved by the corrector alone and reducing the error of the weakest backbones by up to 27%. These gains are robust across different LLM backends and persist even when applied on top of fine-tuned backbones.