CastClaw preprint reports stronger electricity-price forecasts than 16 baselines

Researchers report that CastClaw, an interactive forecasting system, achieved the lowest point-estimate mean squared error and mean absolute error among 16 baselines across five public electricity-price datasets. The preprint describes an AI agent that reads tasks and evidence, selects permitted tools, and requests human feedback. Users define and refine tasks in natural language through command-line and Web interfaces. The system checks results and manages execution limits, historical validation and task state. An npm distribution supports installation, interactive execution and forecast exports. Its North China electricity-load demonstration used offline task execution.