StakeBench - Revealed Side: leaderboard

Revealed-side identification in StakeBench (2026): given a comment by a user who holds a prediction-market position, infer which side that position is on. The score is strict accuracy in percent, abstentions counted wrong, macro-averaged over topic and platform splits; higher is better.

Metric: Strict accuracy (%) on G2: which side (YES or NO) a positioned commenter holds, abstaining counted as wrong, Polymarket and Manifold comments of resolved markets, macro-average over 18 topic and platform splits, greedy decoding; higher is better. Source: arxiv.org. Saturation forecast: Around 2031. 15 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Llama 3.2 3B Instruct53.4#7241
2Fino1-8B52.4#6784
3FinMA-30B47.6
4GPT-5.544.8#54
5Fin-Chat-7B44.5
6Qwen 3 32BNon-reasoning43#972
7Qwen 3 14BNon-reasoning42.3#1269
8Gemma 2 9B39.7#3201
9Gemini 2.5 Flash38.4#463
10Claude Haiku 4.538.1#583
11Fino1-14B36.3
12DeepSeek R1 Distill Llama 8B34.5#7059
13Qwen 3 30B A3BNon-reasoning34.4#1156
14Mistral 7B Instruct (v0.3)33.5#6750
15Qwen 3 8BNon-reasoning32.5#1873

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

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