BacktestBench - Metrics Calculation: leaderboard

The Metrics Calculation task family of BacktestBench (2026): computing a backtest metric such as return or drawdown from a natural-language strategy, from a strategy written in natural language over a database of real market records. Each model drives the same retrieval, SQL and Python-execution pipeline; the score is accuracy in percent against the executed reference backtest; higher is better.

Metric: Back-test accuracy (%) on the Metrics Calculation task family (computing a backtest metric such as return or drawdown from a natural-language strategy), synthetic BacktestBench test split (the expert-crafted subset excluded), each model driving the AutoBacktest pipeline of the authors (Summarizer, SQL Retriever and Python Coder agents) over a PostgreSQL market database, temperature 0.6; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 23 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Gemini 3 Pro51.67#145
2Qwen 3 Max44.14#415
3Seed 1.843.22#252
4GLM-4.739.13#380
5Qwen 3 235B A22B 2507 (Thinking)34.71#466
6GPT-OSS-120B30.86#704
7Kimi K2 (Thinking)30.02
8DeepSeek V3.229.18#419
9Qwen 3 30B A3B 2507 (Thinking)28.67#764
10Qwen 3 32B27.93#972
11Qwen 3 Next 80B A3B (Thinking)26.02#594
12MiniMax-M2.121.98#575
13Qwen 3 Coder Plus20.59#720
14MiMo-V2-Flash17.43#684
15GPT-OSS-20B17.33#1183

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=backtestbench-metrics-calculation · How It Works · Data refreshed daily, snapshot 2026-11-21.