TSorchestra

[Ongoing Project]

Time Series Orchestra (TSorchestra) is a novel ensemble framework designed for zero-shot time series forecasting. It's a curated collection of time series foundation models (TSFMs) that leverages each TSFM's strengths to create something greater than the sum of its parts, yielding SOTA performance.

Set Up

---

1. Create a new conda environment named tso from our .yml file:

conda env create -f environment.yml

2. Download the GIFT-Eval benchmark from Hugging Face:

mkdir data
huggingface-cli download Salesforce/GiftEval --repo-type=dataset --local-dir data

3. Set up the environment variable for loading the datasets: bash echo "GIFT_EVAL=data" >> .env

Usage

---

Run our evaluation script to reproduce our results:

chmod +x ./cli/eval.sh
./cli/eval.sh

Cite

@article{cao2025conversational,
  title={Conversational Time Series Foundation Models: Towards Explainable and Effective Forecasting},
  author={Cao, Defu and Gee, Michael and Liu, Jinbo and Wang, Hengxuan and Yang, Wei and Wang, Rui and Liu, Yan},
  journal={arXiv preprint arXiv:2512.16022},
  year={2025}
}