Unified ranking

Time Series Classification Leaderboard

Compare foundation models in the TSC-FM Benchmark by accuracy, average rank and coverage, with Standard and Low-shot evaluation.

Only configurations covering all 220 unique datasets are eligible for the Standard Overall leaderboard.

Standard Overall Average Accuracy is the unweighted mean across 220 unique datasets. Every dataset receives equal weight. Other views use their displayed dataset coverage.

Standard and Low-shot specify the label budget; ICL and Frozen-Probe filter the Standard Overall results by prediction approach.

Verified Standard
1

GDUT · Huawei Noah’s Ark Lab · Université Paris Cité · MBZUAI · Shantou University

78.57
2

Huawei Noah’s Ark Lab · Université Paris Cité

78.41
3

Huawei Noah’s Ark Lab · Université Paris Cité

78.14
4

Huawei Noah’s Ark Lab · Université Paris Cité

78.09
5

Huawei Noah’s Ark Lab · Université Paris Cité

77.52
6

SCUT · A*STAR · Khalifa University

76.66
7

Huawei Noah’s Ark Lab · Université Paris Cité

76.51
8

Huawei Noah’s Ark Lab · Université Paris Cité

76.47
9

Auton Lab, Carnegie Mellon University

75.85
10

Inria · École Normale Supérieure–PSL

75.64
11

Auton Lab, Carnegie Mellon University

75.14
12

Huawei Noah’s Ark Lab · Université Paris Cité

74.06
13

Inria · École Normale Supérieure–PSL

73.85
14

Auton Lab, Carnegie Mellon University

71.89
15

Microsoft Research · Peking University

66.59

17 methods

Page 1 of 2

Server-rendered result summary

Published leaderboard summaries

These tables summarize the same verified results as the interactive explorer above and are included directly in the page HTML.

Protocol v0.1 · Release 4377b707c842

Standard Overall

17 complete configurations across 220 datasets. Each dataset has equal weight; rank is ordered by mean accuracy and average rank is recomputed within this view.

Top five configurations for Standard Overall
RankConfigurationAvg. AccuracyAvg. RankCoverage
1ChorusTIC78.575.68220 / 220
2Mantis+ScaledLR78.416.15220 / 220
3Mantis+LR78.146.73220 / 220
4MantisV2+ScaledLR78.096.50220 / 220
5MantisV2+LR77.526.97220 / 220

5-shot Overall

18 complete configurations across 201 eligible datasets with five labeled examples per class. Results are ranked only against this budget and scope.

Top five configurations for 5-shot Overall
RankConfigurationAvg. AccuracyAvg. RankCoverage
1MantisV2+ScaledLR64.887.22201 / 201
2Mantis+LR64.837.35201 / 201
3Mantis+ScaledLR64.787.34201 / 201
4Mantis+SVM64.767.35201 / 201
5ChorusTIC64.088.12201 / 201

10-shot Overall

18 complete configurations across 178 eligible datasets with ten labeled examples per class. Results are ranked only against this budget and scope.

Top five configurations for 10-shot Overall
RankConfigurationAvg. AccuracyAvg. RankCoverage
1MantisV2+ScaledLR70.396.73178 / 178
2Mantis+SVM70.386.58178 / 178
3Mantis+ScaledLR70.016.90178 / 178
4Mantis+LR69.747.30178 / 178
5ChorusTIC69.197.42178 / 178

Univariate and multivariate leaders

The two scopes use different dataset pools, so compare configurations within the same column rather than comparing their accuracy values as if they came from one test set.

Standard

Univariate · 140 datasets
TimEE leads at 83.30 average accuracy.
Multivariate · 80 datasets
Mantis+LR leads at 74.45 average accuracy.

5-shot

Univariate · 124 datasets
TimEE leads at 69.91 average accuracy.
Multivariate · 77 datasets
Mantis+LR leads at 62.47 average accuracy.

10-shot

Univariate · 108 datasets
TimEE leads at 75.14 average accuracy.
Multivariate · 70 datasets
Mantis+SVM leads at 67.52 average accuracy.

How ranking works. Average Accuracy is the unweighted mean across the eligible datasets shown by Coverage. Exact accuracy ties within 1e-12 receive average ranks. Low-shot budgets and modality scopes remain separate.

Read the complete ranking rules and reproducibility protocol. The release identifier above is a content fingerprint for the currently published public results.

How to compare time series classification foundation models

Start with Standard Overall to compare configurations covering all 220 datasets. Select Univariate or Multivariate when that matches your task. A model name can appear with several classifier heads: the head is part of the evaluated configuration, so the rankings compare complete prediction procedures.

Some models cover a narrower scope. TiCT and TimEE have results on all 140 univariate datasets under Standard evaluation. Select Standard Univariate to compare them with other models; neither has published Overall or Low-shot results here.

For few-shot time series classification, select Low-shot and then 5-shot or 10-shot. Compare models within the same budget and scope; dataset eligibility changes with the label budget. Average accuracy weights each eligible dataset equally, while average rank describes performance relative to the comparison set.

The leaderboard reports classification, where a sequence is assigned a class label. It does not rank time series forecasting models on future-value prediction. Inspect the model pages for checkpoints and configuration results before selecting a model.