USTC-AGI/CastStar

This is a benchmarking artifact for the GIFT-Eval submission. CastStar is an agentic time-series forecasting system built on top of multiple foundation forecasting models. The submitted GIFT-Eval result uses pre-computed forecasts and a meta-selection/ensemble layer over the model pool. It is intended to make the leaderboard submission reproducible and comparable under the GIFT-Eval protocol.

What This Is

CastStar combines forecasts from a pool of time-series foundation models and uses a learned gating strategy to select or weight base-model predictions for each forecasting setting. The system follows the same general meta-forecasting idea as FFORMA-style model selection, where lightweight time-series features and dataset metadata guide the choice of forecasting experts.

The GIFT-Eval submission evaluates CastStar on all 97 dataset configurations using the official CSV format.

Model Pool

CastStar uses the following base forecasting models. Compared with the Toto-2.0-FnF setup, CastStar only uses the Toto-2.0-2.5B checkpoint from the Toto family.

#	Model	Family
0	chronos-2	Chronos
1	timesfm-2.5	TimesFM
2	flowstate	FlowState
3	tirex	TiRex
4	patchtst-fm	PatchTST
5	toto-2.0-2.5b	Toto 2.0

Key Features

• Agentic forecasting system: CastStar uses multiple forecasting experts and a decision layer to produce final probabilistic forecasts.
• Foundation-model pool: The model pool includes Chronos-2, TimesFM-2.5, FlowState, TiRex, PatchTST-FM, and Toto-2.0-2.5B.
• GIFT-Eval compatible: Results are submitted in the official GIFT-Eval format with 97/97 dataset configurations and 11 evaluation metrics.
• No test-data leakage: The submitted configuration reports no GIFT-Eval test-data leakage.
• Benchmark-focused artifact: The submitted result is designed for leaderboard evaluation and reproducibility under the GIFT-Eval protocol.

GIFT-Eval Submission

The submitted files are:

results/CastStar/all_results.csv
results/CastStar/config.json

Submission metadata:

{
  "model": "CastStar",
  "model_type": "agentic",
  "model_dtype": "float32",
  "model_link": "https://huggingface.co/USTC-AGI/CastStar",
  "code_link": "https://github.com/ustc-time-series/CastStar",
  "org": "CastStar",
  "testdata_leakage": "No",
  "replication_code_available": "No"
}

Additional Resources

• GIFT-Eval benchmark
• GIFT-Eval GitHub repository
• CastStar GitHub repository
• CastStar model page

Citation

If you use CastStar, please cite the corresponding CastStar paper or repository when available.