RTA-Bench1 - 2026 Blind - Sharpe: leaderboard

RTA-Bench1 - 2026 Blind - Sharpe — RTA-Bench1 static67-entry/six-model release; blind ticker condition, full native period windows2020/2022/2026 remain separate. Mean over native1or3seeds; five-session decisions,56megacap universe,long-only,max10positions,base5bps slippage plus native realistic gap/impact/volume cost rules,news enabled. These are retrospective trading decisions: historical windows may be remembered;2026also includes pre-cutoff observations. Full-window return/Sharpe/signed drawdown are native observations, not after-cutoff headline values or live investment results. Model-specific cutoffs,seed counts,windows,all raw row statistics and source file pointers remain provenance. Historical Sonnet4.6/Haiku5.5 blind implementation could expose dates/dollar amounts in headlines; other four were evaluated after masking changes. Native entry package commits, per-seed model.call and access are retained. Llama is explicitly3B via Ollama16384context, but exact checkpoint/quantization is unreported. Rule/market anchors and named-minus-blind contamination diagnostics are not scored model rows. Metric: Annualized Sharpe.

Metric: Annualized Sharpe. Source: raw.githubusercontent.com. Saturation forecast: Around September 2027. 6 models tracked. Updated 15 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.6 LunaNon-reasoning1.1#122
2Claude Sonnet 4.6Non-reasoning0.4431#165
3Claude Haiku 4.5Non-reasoning0.4161#575
4Gemini 3.5 FlashMinimal0.3257#100
5Claude Haiku 5.5Non-reasoning-0.05977#70
6Llama3.2 3BRTA run: Ollama; checkpoint and quantization unreported-0.3156

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=rta-bench1-2026-blind-sharpe · How It Works · Data refreshed daily, snapshot 2026-11-16.