Tag
The paper introduces Crafter, an agent for corrective feature discovery that mines the residual of frozen black-box forecasters using compositional search and LLM-generated features, achieving up to 27% error reduction across six datasets and backbones.