MetaCaster: Meta-Harness-Optimized Agent for End-to-End Few-Shot Learning of Lightweight Time Series Forecasters
The paper proposes using agents to build small task-specific forecasters from only a few examples, rather than using the agents as forecasters themselves.
MetaCaster combines agentic data generation with textual context to train lightweight time-series models for scarce-data settings. The authors frame this as a way around the cost of foundation models and the data appetite of compact forecasters. They report tests across 18 datasets, 23 lightweight forecasters, and 14 baselines, claiming gains in data and compute efficiency while preserving forecast quality. The paper is listed as accepted by EMNLP 2026. ArXiv · AI/CL/LG's note
MetaCaster combines agentic data generation with textual context to train lightweight time-series models for scarce-data settings. The authors frame this as a way around the cost of foundation models and the data appetite of compact forecasters. They report tests across 18 datasets, 23 lightweight forecasters, and 14 baselines, claiming gains in data and compute efficiency while preserving forecast quality. The paper is listed as accepted by EMNLP 2026. ArXiv · AI/CL/LG's note
score 4