Imagine a forecaster that you cannot touch: retraining it would cost too much in money and time, or the regulatory re-validation of a frozen model is a bureaucratic nightmare. CRAFTER (Corrective Residual Agent with Feature-based Temporal Exploration and Reasoning) answers with an elegant workaround: it does not modify the frozen backbone but focuses on the residual, i.e., the systematic errors that the model makes. The result is lightweight, interpretable, and surprisingly effective corrections.

CRAFTER employs two complementary generators. A compositional search digs into raw input channels to find useful combinations; the other generator queries a large language model (LLM) and asks it to invent features with explicit names, binary flags, or short executable code snippets. Every candidate, regardless of origin, passes through a validation-grounded gate: if the proposed feature does not improve performance, it is rejected. A final corrector applies the accepted features or leaves the original forecast unchanged.

Decoupling backbone and corrector is the key. Across six public datasets and six frozen backbones, CRAFTER surpasses every dedicated feature-engineering system at any feature budget, roughly doubling the improvement gained by a feature-less corrector and reducing the error of the weakest backbones by up to 27%. These gains hold even when the system is applied on top of models that have already been fine-tuned, and they are robust across different LLM backends.

The significance goes beyond the numbers. For those operating in regulated or on-premise environments, where retraining a model can mean months of audits and compliance costs, the ability to update a forecaster without touching its core is a strategic lever. The corrective features are interpretable by design: names, flags, and code snippets make every corrector decision traceable, easing explainability toward regulators and internal stakeholders. The LLM, while a “creative” generator, is kept in check by the validation gate, which prevents the injection of spurious correlations; and because the method is agnostic to the linguistic backend, everything can run on a local LLM and keep data within the corporate perimeter.

There is also a structural signal for the industry. CRAFTER turns corrective feature discovery into a systematic and measurable process, so much so that the authors propose it as a scientific instrument to attribute forecast improvements solely to the feature source. In a discipline often dominated by the race to ever-larger models, the paper reminds us that a post-hoc, interpretable corrector can deliver a better return on investment than fine-tuning, shifting the focus from training hardware to inference and monitoring pipelines. For those designing local AI infrastructure, this means being able to evolve forecasting systems with a contained impact on GPU resources, energy, and compliance. It is the kind of pragmatism that could make “corrective feature discovery” a permanent piece of the MLOps lexicon, much like hyperparameter tuning is today.