Skip to content

Overall Benchmark Arena

Cumulative Elo leaderboard combining all 56 benchmark datasets (18 Regression on test RMSE + 38 Classification on test Accuracy).
Click any column header to sort. Evaluates models across Rust, Python, and R under identical 300s wall-clock isolation.

Arena Navigation:
- Regression Arena (18 datasets)
- Classification Arena (38 datasets)
- Datasets Metadata & Sources


Executive Summary & Cross-Task Insights

1. Unified Arena Overview

The cumulative TabArena-v0.1 arena provides an empirical tournament across 56 diverse datasets spanning physical sciences, medical diagnosis, financial churn, industrial failures, and web analytics. Every algorithm was tested under strict serial evaluation with a 300-second wall-clock timeout per run.

2. High-Altitude Rankings

  1. grf_llf (Rank 1, Elo 1585.4): Leads in regression via Generalized Random Forests with local linear leaves.
  2. CatBoost (Rank 2, Overall Elo 1477.5): The most consistent dual-domain performer across all 56 datasets (Reg: 1479.0, Cla: 1388.0).
  3. perpetual (Rust) (Rank 3, Overall Elo 1391.8): High-speed pure-Rust gradient booster with automatic stopping and zero hyperparameters.
  4. sklearn RF & ChimeraBoost (Ranks 4-5, Elo 1368.5 & 1354.9): Random forests remain extraordinarily resilient across diverse tabular domains without manual tuning.
  5. EBM (Explainable Boosting Machine) (Rank 6, Elo 1353.5): Highest-ranking fully interpretable model, though computationally expensive (~44.7s average fit time).
  6. ModelTree(all) (Rank 7, Overall Elo 1333.9): The top-performing pure-Rust model tree architecture. It outperforms traditional trees (ctree, rpart, sklearn tree) by 300+ Elo points and matches or exceeds standard gradient boosters on many tasks by substituting flat leaf constants with regularized local linear models.

3. Language & Runtime Economy

  • Rust Core Advantage: ModelTree and perpetual provide native memory safety and multi-threaded throughput with zero reliance on external BLAS/Fortran runtimes.
  • Microsecond Leaf Updates: While full GBDT re-fitting requires dozens of seconds, ModelTree enables frozen-topology leaf retraining in < 2 ms, allowing high-throughput online serving on streaming data.

Overall Leaderboard (56 Datasets)

RANK OVERALL ELO MODEL REG ELO CLA ELO DATASETS FIT TIME (±sd) LANG CAT HANDLING
11585.4grf_llf1585.4·16/5623.17±31.44sRnone
21477.5CatBoost(200)1479.01388.056/565.59±19.65sPythoninternal
31391.8perpetual (Rust)1490.21302.341/5615.04±30.44sRustexternal
41368.5sklearn RF1670.41279.056/569.76±22.40sPythonexternal
51354.9ChimeraBoost1723.51265.356/567.42±11.86sPythoninternal
61353.5EBM1362.81264.044/5644.66±49.61sPythoninternal
71333.9ModelTree(all)1123.31244.354/5617.13±36.34sRustinternal
81318.3bart1237.21228.836/5619.07±29.65sRinternal
91293.5ranger1685.41204.054/5618.10±30.97sRinternal
101281.7sklearn GBM1120.61192.255/568.77±19.04sPythonexternal
111277.4c50_rules·1187.833/563.99±5.63sRinternal
121260.3HistGradientBoosting1411.71170.850/561.34±3.09sPythoninternal
131248.9c50_tree·1159.436/565.00±7.76sRinternal
141237.7LightGBM(200)1525.91148.256/560.60±0.80sPythoninternal
151200.4CTBoost1283.21110.946/560.65±1.56sPythoninternal
161165.8ExtraTrees1623.41076.256/565.77±19.44sPythonexternal
171126.4forust (Rust)1363.61036.948/561.19±0.81sRustexternal
181108.3MARS (mars-earth)885.31018.818/5627.24±40.59sPythonexternal
191066.7lmtree1066.7·14/5620.22±47.48sRinternal
201065.6modeltrees lmtree1065.6·14/5615.29±51.54sPythoninternal
211037.7ctree668.6948.144/569.71±37.27sRinternal
221036.6ModelTree(opt)852.0947.156/562.37±8.95sRustinternal
231015.2ModelTree(path)595.9925.656/568.23±31.06sRustinternal
241013.9ModelTree(mean)506.5924.456/566.11±30.50sRustinternal
25999.1grf_rf1242.5909.527/5615.45±20.17sRnone
26964.1cubist964.1·17/566.23±15.19sRinternal
27925.7ModelTree(winnow)980.2836.154/566.54±10.19sRustinternal
28889.7sklearn tree562.4800.156/560.85±3.85sPythonexternal
29880.5KNN679.5791.056/560.26±0.53sPythonexternal
30849.9rpart335.1760.453/562.02±5.98sRinternal
31827.3modeltrees glmtree·737.716/5623.99±39.03sPythoninternal
32816.8statsmodels MNLogit·727.65/560.03±0.01sPythonnone
33814.9ModelTree(v0.3.3-default)814.9·18/560.11±0.35sRustinternal
34790.9mboost859.2701.334/5616.11±37.99sRinternal
35790.7modeltrees multinomtree·701.115/5611.53±25.66sPythoninternal
36777.4glmtree1033.1687.930/5628.38±48.39sRinternal
37767.4mltreelib MARS(deg=2)767.4·17/563.19±7.11sPythonunknown
38756.3earth654.5666.836/565.92±19.25sRinternal
39748.3sklearn LogReg·658.738/564.27±6.32sPythonexternal
40701.7statsmodels OLS701.7·8/560.33±0.68sPythonnone
41686.9ModelTree(leaf=mars,deg=1)686.9·17/560.30±0.47sRustinternal
42677.4ModelTree(leaf=mars,deg=2)677.4·17/560.74±1.38sRustinternal
43656.8mltreelib MARS(deg=1)656.8·17/560.75±1.30sPythonunknown
44583.4trust-free583.4·4/5614.37±7.33sPythonnone
45554.9sklearn Linear554.9·18/560.09±0.23sPythonexternal
46549.6statsmodels RLM549.6·8/5617.48±43.02sPythonnone
47370.2statsmodels QuantReg370.2·7/5622.52±59.15sPythonnone