New AI agent CRAFTER fixes flawed forecasts automatically
New AI tool CRAFTER doubles forecasting accuracy by mining model flaws without fine-tuning
A team led by Fangxin Wang at the University of Illinois Chicago has unveiled CRAFTER (Corrective Residual Agent with Feature-based Temporal Exploration and Reasoning), a novel agent designed to automatically discover corrective features for black-box AI forecasters without fine-tuning.
CRAFTER operates by analyzing the residual errors of frozen models and mining interpretable features that capture recurring failure patterns. The system combines compositional search over raw input channels with an LLM that proposes named feature combinations, binary flags, and executable code. A validation-grounded gate filters candidates, and a lightweight corrector applies only the most effective features. In experiments spanning six public datasets and six frozen backbones, CRAFTER consistently outperformed all dedicated feature-engineering baselines, roughly doubling the improvement achieved by correctors alone and cutting error rates by up to 27%.
- CRAFTER automatically discovers corrective features for flawed AI forecasters without fine-tuning
- Combines compositional search with LLM-generated feature proposals and a validation gate
- Cut forecast errors by up to 27% across 6 datasets and 6 models, doubling corrector performance
Why It Matters
CRAFTER offers a scalable way to fix AI forecasting systems without costly retraining, unlocking higher accuracy for enterprises relying on black-box models.