TSC-FM Benchmark is a time series classification benchmark for foundation models, spanning 198 unique datasets. Compare Standard and Low-shot results across in-context and frozen-probe approaches, with equal weighting for every dataset.
Total
198
Univariate
128
Multivariate
70
Tracks
4
198 unique datasets
Ranked by mean accuracy with equal weighting across 198 unique datasets.
1 | GDUT · Huawei Noah’s Ark Lab · Université Paris Cité · MBZUAI · Shantou University | 78.56 |
2 | Huawei Noah’s Ark Lab · Université Paris Cité | 77.73 |
3 | SCUT · A*STAR · Khalifa University | 76.87 |
4 | Huawei Noah’s Ark Lab · Université Paris Cité | 76.62 |
5 | Auton Lab, Carnegie Mellon University | 76.23 |
Benchmark purpose
A pretrained backbone is only one part of a classifier. In-context methods use labeled examples at prediction time, while frozen encoders need a classifier fitted on their representations. Changing that classifier or the label budget changes the configuration being evaluated. A model name alone does not describe this comparison.
TSC-FM makes the configuration, label budget and dataset coverage explicit. It compares classification predictions on official train/test splits, with equal weight for each dataset. Researchers can check which configurations lead within a setting, how the ordering differs by input type, and where coverage limits a comparison.
These observations use the same published results as the leaderboard. They describe the evaluated configurations and task pools; they are not claims of universal superiority or statistical significance.
ChorusTIC leads 12 complete configurations with 78.56% mean accuracy and 3.94 average rank across 198 / 198 datasets. MantisV2+LR follows at 77.73%. The mean-accuracy gap is 0.83 percentage points; this aggregate gap alone does not establish statistical significance.
Compare Standard resultsAt five labels per class, MantisV2+ScaledLR leads with 66.33% mean accuracy on 140 / 140 datasets. At ten labels per class, Mantis+SVM leads with 72.07% on 121 / 121 datasets.
The eligible pools differ, so subtracting these means would not isolate the effect of adding labels.
Check low-shot eligibilityChoose the relevant input type before using an Overall rank to shortlist configurations.
Compare modality resultsStart with the input type and the number of labels available for your task. Standard uses the full official training split; 5-shot and 10-shot use five or ten labels per class on their own eligible dataset pools. Then choose whether you need in-context prediction or a frozen encoder with a fitted classifier.
The unit of comparison is a complete classification configuration: a pretrained model, its target prediction approach and a defined label budget. The same backbone can appear with different classifier heads, so compare configurations within the same dataset scope and evaluation setting rather than treating a model name as a single score.
Use the classification leaderboard to compare average accuracy, average rank and dataset coverage. The catalog of foundation models for time series classification provides evaluated configurations and available paper, code and checkpoint references, while the evaluation protocol explains full-training and 5-shot / 10-shot comparisons. Low-shot coverage is reported separately for each label budget.
These rankings summarize classification accuracy on the published task pools. They do not establish the best model for every individual dataset, measure forecasting quality, or rank deployment latency and memory. The public site reports aggregate results and dataset profiles; it does not publish individual dataset or suite identities. Read the ranking definitions and dataset coverage before transferring a result to a new application.
Research behind the benchmark
The accompanying survey introduces the TSC-FM Benchmark and examines how the target classification procedure affects comparisons between foundation models. Use the live classification results alongside the paper and the evaluation protocol to check the setting, configuration and dataset scope of a comparison.
Preprint · Version 1 · Not peer-reviewed · Published September 24, 2026
Jointly released by DMIR Lab and Huawei Noah’s Ark Lab. Maintained by Juntao Fang. New model submissions are reviewed before their results enter the official leaderboard.
Read the survey and benchmark paperProtocol
Standard and Low-shot define the label budget. ICL and Frozen-Probe describe how predictions are produced; these approaches can appear in either budget setting. Results remain separated by budget and dataset scope.
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.
