Aggregate analysis

Capability Map

Explore a fixed full-coverage core through separate Evaluation Paradigm and Model Origin taxonomies, with backbone- and dataset-equal aggregation.

Fixed Overall capability core

The analysis uses the same complete method universe as the Standard Overall leaderboard: 17 configurations across 8 backbones, each evaluated on all 220 datasets.

Configuration scores are averaged within each backbone first. Backbones and datasets then receive equal weight, preventing classifier-rich backbones from receiving extra influence.

220

datasets

17

configurations

8

backbones

220 / 220

result coverage

Comparison taxonomy

Choose one independent analysis lens

Groups are mutually exclusive within each taxonomy. Evaluation Paradigm and Model Origin are calculated and ranked separately.

Property coverage

Choose a benchmark attribute

Every property has reviewed metadata for all 220 datasets; coverage remains visible for audit transparency.

Capability heatmap

Evaluation Paradigm × modality

220 of 220 datasets contribute to this property view. N is the dataset count in each condition.

Fixed core · 17 configurations · 8 backbones
Taxonomy groupUnivariate1 channelMultivariate2 or more channels
Native time-series ICL1 backbone · 1 configuration

81.0%

N=140

74.4%

N=80

Frozen representation + classifier5 backbones · 14 configurations

74.9%

N=140

69.8%

N=80

Tabular ICL adaptation2 backbones · 2 configurations

77.5%

N=140

69.9%

N=80

Aggregate evidence

Condition-level detail

Every row is a fixed-core aggregate; no dataset-level record is exposed or downloadable.

Hierarchical Accuracy

For each dataset, configuration accuracies are averaged within their shared backbone. Backbone scores are then averaged with equal backbone weight within each taxonomy group. The resulting group scores are averaged with equal dataset weight within the selected condition.

Best / Tied Best / Other

Best / Tied Best / Other counts whether a group is uniquely highest / tied for highest / below the highest within the active taxonomy on each dataset.

ConditionTaxonomy groupAccuracyAvg. Group RankBest / Tied / OtherDatasetsBackbonesConfigurations
UnivariateNative time-series ICL81.0%1.4978 / 2 / 6014011
UnivariateFrozen representation + classifier74.9%2.597 / 0 / 133140514
UnivariateTabular ICL adaptation77.5%1.9253 / 2 / 8514022
MultivariateNative time-series ICL74.4%1.5246 / 2 / 328011
MultivariateFrozen representation + classifier69.8%2.428 / 1 / 7180514
MultivariateTabular ICL adaptation69.9%2.0624 / 2 / 548022

Reading guide

Methodology and privacy boundaries

Fixed result universe

The fixed core matches the Standard Overall leaderboard: 17 configurations across 8 backbones, each with complete 220 / 220 coverage.

Rank definition

Average Rank ranks the mutually exclusive groups within the active taxonomy on each dataset, using average ranks for exact ties.

Separate taxonomies

Evaluation Paradigm and Model Origin are separate taxonomies. Groups are mutually exclusive within each taxonomy and are never ranked across taxonomies.

Public data boundary

Only aggregate condition-level statistics are published. Dataset identities, source mappings, backbone names, and configuration names are excluded.

Models included in this analysis

The heatmap aggregates 17 configurations from these 8 backbones, each with results on all 220 datasets. Their model pages list the individual scores and classifier settings.

The metric definitions explain accuracy and win counts. The aggregation method describes how classifier results contribute to a backbone score and then to a group score.

Average Group Rank compares groups within a taxonomy. For individual configurations, use the Standard Overall leaderboard and its ranking rules. The dataset profile gives the composition of the 220-dataset benchmark.