CastStar

CastStar is an agentic time-series forecasting system built on multiple forecasting foundation models. This repository contains the frozen inference assets for reproducing the CastStar submission on all 97 GIFT-Eval dataset configurations.

Model Pool

Model	Source
Chronos-2	amazon/chronos-2
TimesFM-2.5	google/timesfm-2.5-200m-pytorch
FlowState	ibm-research/flowstate
TiReX	NX-AI/TiRex-1.1-gifteval
PatchTST-FM	ibm-research/patchtst-fm-r1
Toto-2.0-2.5B	Datadog/Toto-2.0-2.5B

The original model repositories retain their respective licenses and terms.

Artifact Layout

CastStar-HF/
├── README.md
├── manifest.json
├── calibration/
│   └── bucket_class_biases.csv
├── models/
│   ├── pre/
│   └── post/
└── test/
    ├── input_features/
    ├── predictions/
    └── predicted_features/

The bundle contains only the frozen files required by the reproduction notebook. It does not contain training data or training code.

Reproduction

1. Clone and install GIFT-Eval.
2. Download the Salesforce/GiftEval dataset.
3. Open notebooks/caststar.ipynb.
4. Set the local GIFT-Eval dataset path and run the notebook.

The notebook downloads this repository, runs frozen CastStar inference, evaluates the forecasts with the official GIFT-Eval metrics, and writes a submission-format CSV.

GIFT-Eval Submission

{
  "model": "CastStar",
  "model_type": "agentic",
  "model_dtype": "float32",
  "model_link": "https://huggingface.co/USTC-AGI/CastStar",
  "code_link": "https://github.com/SalesforceAIResearch/gift-eval/blob/main/notebooks/caststar.ipynb",
  "org": "USTC-AGI",
  "testdata_leakage": "No",
  "replication_code_available": "Yes"
}

Intended Use

These assets are intended for reproducing and inspecting the CastStar GIFT-Eval benchmark submission.