CastStar is a time-series forecasting agent.
On GIFT-Eval, the current experimental configuration achieves:
| Metric | Result |
|---|---|
| Normalized MASE | 0.672452 |
| Normalized CRPS | 0.459657 |
| Composite | 0.566054 |
[2026-07-02] The current result ranks first by MASE Rank and second by CRPS Rank among the evaluated submissions at the time of evaluation.
The training, data-preparation, and inference code is being organized and documented. The complete implementation and reproduction instructions will be open-sourced soon.
@misc{caststar2026,
title = {CastStar},
author = {{USTC-AGI Group}},
year = {2026},
howpublished = {\url{https://github.com/ustc-time-series/CastStar}},
note = {GitHub repository},
}