PortBench-QA: leaderboard
PortBench-QA is the static question-answering layer of PortBench (Zhao, Chen and Su, 2026): portfolio-management questions generated from 2015-2025 market data with answers derived by formula or optimizer, spanning return-direction forecasts, value at risk, position sizing, pairwise and maximum-Sharpe allocation, rebalancing and regime-driven allocation. Mean item score in percent over seven templates of 50 test-period questions each, zero-shot at temperature 0 with thinking on; higher is better.
Metric: Mean item score (%; mean of seven QA templates, 50 test questions per model per template, thinking on, temperature 0). Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.7 Max (Thinking) | 81.9 | #142 (Qwen 3.7 Max) |
| 2 | Qwen 3.6 Plus (Thinking) | 78.3 | #217 (Qwen 3.6 Plus) |
| 3 | GLM-5.1 (Thinking) | 75.7 | #175 (GLM-5.1) |
| 4 | Qwen 3.6 35B A3B (Thinking) | 68.4 | #461 (Qwen 3.6 35B A3B) |
| 5 | Kimi K2.6 (Thinking) | 51.1 | #173 (Kimi K2.6) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=portbench-qa · How It Works · Data refreshed daily, snapshot 2026-11-12.