ChorusTIC

ChorusTIC is a classification-native foundation model for training-free univariate and multivariate time series classification. Given labeled context examples, ChorusTIC directly predicts query labels through in-context learning, without fitting a target-specific classifier or updating model parameters.

🏆 #1 on the TSC-FM Standard Overall Leaderboard, achieving an Overall Average Accuracy of 78.56 across all 198 time series classification datasets.

📄 Paper · 💻 Code · 🤗 Checkpoint · 🏆 TSC-FM Leaderboard · 📊 Benchmark Results

---

Model Overview

ChorusTIC is designed for support-conditioned time series classification, where predictions are made directly from a labeled context set and a set of query samples.

The model combines two complementary levels of modeling:

• Signal-level Chorus, which uses Random Subchannel Slot Concatenation (RSSC) together with a shared dual-axis time series encoder to capture temporal patterns and cross-channel interactions under heterogeneous channel configurations.
• Task-level Chorus, which performs support-conditioned classification through Column Distribution Modeling (CDM), Row-wise Feature Interaction, and In-Context Learning (ICL).

This design allows a single pretrained model to perform both univariate and multivariate time series classification without task-specific parameter optimization at inference time.

Key Characteristics

• Training-free inference: no fine-tuning or target-specific classifier fitting.
• In-context classification: predictions are conditioned directly on labeled support examples.
• Unified univariate and multivariate modeling: the same pretrained model handles datasets with different channel configurations.
• Support-conditioned prediction: the model adapts its predictions to the target task through the provided context set rather than parameter updates.
• Foundation-model evaluation: evaluated across a large and heterogeneous collection of time series classification datasets.

---

Benchmark Results

ChorusTIC is evaluated on the TSC-FM Time Series Classification Benchmark, a unified benchmark covering 198 unique datasets across both univariate and multivariate time series classification.

On the Standard Overall evaluation, ChorusTIC achieves:

Metric	Result
TSC-FM Standard Overall Rank	#1
Overall Average Accuracy	78.56
Dataset Coverage	198 / 198
Inference Paradigm	Training-free In-Context Learning

The TSC-FM Overall score is computed with equal weighting across the evaluated datasets.

For complete evaluation details, configurations, and dataset-level results, see:

• ChorusTIC benchmark results
• TSC-FM Leaderboard
• TSC-FM Evaluation Protocol
• TSC-FM Benchmark Homepage

Leaderboard rankings may evolve as new models and evaluation results are added. Please refer to the official TSC-FM leaderboard for the latest ranking.

---

Released Checkpoint

This repository provides the pretrained ChorusTIC checkpoint and its corresponding model configuration:

ChorusTIC.ckpt
model_hparams_latest.json

The checkpoint can be downloaded with the Hugging Face CLI:

pip install -U huggingface_hub

hf download DMIRLAB/ChorusTIC \
  --local-dir Checkpoints_ChorusTIC

The resulting directory should have the following structure:

Checkpoints_ChorusTIC/
├── ChorusTIC.ckpt
└── model_hparams_latest.json

Model loading, preprocessing, inference, and UCR/UEA evaluation code are provided in the official ChorusTIC GitHub repository.

---

Intended Use

ChorusTIC is intended primarily for research on time series foundation models and training-free classification.

Typical use cases include:

• Training-free time series classification.
• In-context classification with labeled support examples.
• Univariate time series classification.
• Multivariate time series classification.
• Evaluation of foundation models on heterogeneous time series datasets.
• Reproduction and extension of the experiments presented in the accompanying paper.

---

Inputs and Inference Setting

ChorusTIC operates under an in-context classification setting.

For a target classification task, the model receives:

1. A set of labeled context examples, which defines the target task and its class structure.
2. One or more unlabeled query examples to be classified.

The model predicts the labels of the query examples directly from the provided context, without updating the pretrained model parameters or fitting an additional task-specific classifier.

The quality and composition of the context set can therefore affect prediction performance.

---

Resources

• Paper: ChorusTIC: Training-Free Multivariate Time Series Classification via Chorus In-Context Learning
• Official Code: DMIRLAB-Group/ChorusTIC
• Model Checkpoint: DMIRLAB/ChorusTIC
• TSC-FM Benchmark: tsc-fm.dmirlab.com
• Benchmark Results: ChorusTIC on TSC-FM

---

Citation

If you use ChorusTIC in your research, please cite:

@article{fang2026chorustic,
  title   = {ChorusTIC: Training-Free Multivariate Time Series Classification via Chorus In-Context Learning},
  author  = {Fang, Juntao and Xie, Shifeng and Cai, Ruichu and Zheng, Shengji and Li, Zijian and Zhang, Keli and Pan, Lujia and Palpanas, Themis and Hao, Zhifeng},
  journal = {arXiv preprint arXiv:2608.24033},
  year    = {2026}
}

---

License

ChorusTIC is released under the Apache License 2.0.