TSC-FM Benchmark

Benchmarking Foundation Models for Time Series Classification

A unified benchmark spanning 198 unique time series classification datasets, four evaluation tracks, and comparable dataset-level reporting.

198 datasets

Univariate

Multivariate

Low-shot

Training-Free / ICL

Frozen-Probe

Benchmark composition

Total

198

Univariate

128

Multivariate

Tracks

Univariate 64.6%Multivariate 35.4%

Equal dataset weighting

Reviewed metadata coverage

Primary metric: Accuracy

198 unique datasets

Verified Standard

Overall Leaderboard Preview

Ranked by mean accuracy with equal weighting across 198 unique datasets.

				Univariate	Multivariate	Coverage	Evaluation Type
1	ChorusTIC GDUT · Huawei Noah’s Ark Lab · Université Paris Cité · MBZUAI · Shantou University	78.56	3.94	81.16	73.78	198 / 198	In-Context
2	MantisV2+LR Huawei Noah’s Ark Lab · Université Paris Cité	77.73	4.64	80.03	73.51	198 / 198	Frozen Probe
3	UniShape+RF SCUT · A*STAR · Khalifa University	76.87	5.08	78.86	73.22	198 / 198	Frozen Probe
4	MantisV2+RF Huawei Noah’s Ark Lab · Université Paris Cité	76.62	5.42	78.79	72.66	198 / 198	Frozen Probe
5	MOMENT+SVM Auton Lab, Carnegie Mellon University	76.23	5.22	77.98	73.04	198 / 198	Frozen Probe

Dataset profile

A high-level view of the benchmark without publishing individual dataset or suite identities.

Total

198

Univariate

128

Multivariate

Tracks

Protocol

Results remain separated by adaptation setting so each score has a clear scientific meaning.

Full official training splits with dataset-level equal weighting.

Matched label budgets across methods and feasible datasets.

Predictions without target-specific parameter updates.

Frozen representations with a standardized lightweight probe.