# PortBench

Multi-asset portfolio management across six asset classes. It pairs a static QA set with a five-stage decision pipeline whose allocation stage is scored on how well it captures cross-asset correlation, under stress windows and three risk profiles.

## Facts

- Maker: AgenticFinLab; The Hong Kong University of Science and Technology (Guangzhou); Yantai Research Institute of Harbin Engineering University
- Year: 2026
- Use case: Financial trading and portfolios
- Size: 6,269 QA items across 7 templates (4 complexity levels); 5-stage pipeline; 183 market series across 6 asset classes, 2015-2025
- Public split: Code on GitHub; dataset as a Hugging Face collection; raw experimental results as a Hugging Face dataset
- Grading: QA: mean score per template. Pipeline: CEPS, the mean of five stage scores normalised to [0,1] minus a 0.1 penalty for drops between stages, plus Sharpe, max drawdown, volatility and return
- Licence: Apache-2.0 (code); MIT (PortBench-RawData results dataset on HF)
- Submission: none
- Products allowed: Not stated
- Latest version: arXiv v6 (2026-10-09)
- What it does not cover: Live execution costs and tool or agent products; the 12-day real-time run is only a preliminary demo
- Last checked: 1 Nov 2026
- Links: paper https://arxiv.org/abs/2605.27887, site https://portbench.github.io/, data https://huggingface.co/collections/AgenticFinLab/portbench, code https://github.com/AgenticFinLab/portbench
- Directory page: https://benchmarks.finance/benchmarks/portbench/
- Leaderboard: https://benchmarks.finance/leaderboards/portbench/

## Results

### PortBench: CEPS (0-1) (higher is better, 10 models)

| Rank | Model | Provider | Score | Metric | Setting | Reported by | Source | URL | Date |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | GLM-5.1 | Zhipu AI | 0.384 | CEPS (0-1) | pipeline, balanced risk profile | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 2 | Qwen3.6-Plus | Qwen | 0.300 | CEPS (0-1) | pipeline, balanced risk profile | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 3 | Qwen3.7-Max | Qwen | 0.292 | CEPS (0-1) | pipeline, balanced risk profile | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 4 | Kimi K2.6 | Moonshot AI | 0.274 | CEPS (0-1) | pipeline, balanced risk profile | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 5 | Qwen3.6-35B-A3B | Qwen | 0.272 | CEPS (0-1) | pipeline, balanced risk profile | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 6 | DS-V4-Pro | DeepSeek | 0.266 | CEPS (0-1) | pipeline, balanced risk profile | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 7 | DB-2.0-Pro | Not stated | 0.263 | CEPS (0-1) | pipeline, balanced risk profile | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 8 | HY3-Preview | Tencent | 0.258 | CEPS (0-1) | pipeline, balanced risk profile | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 9 | DB-2.0-Lite | Not stated | 0.244 | CEPS (0-1) | pipeline, balanced risk profile | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 10 | DS-V4-Flash | DeepSeek | 0.226 | CEPS (0-1) | pipeline, balanced risk profile | Reported by the authors | Paper Table 2 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |

### PortBench: mean QA score (0-1) (higher is better, 10 models)

| Rank | Model | Provider | Score | Metric | Setting | Reported by | Source | URL | Date |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | DS-V4-Flash | DeepSeek | 0.819 | mean QA score (0-1) | static QA, mean over T1-T7 (secondary) | Reported by the authors | Paper Table 1 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 1 | Qwen3.7-Max | Qwen | 0.819 | mean QA score (0-1) | static QA, mean over T1-T7 (secondary) | Reported by the authors | Paper Table 1 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 3 | DS-V4-Pro | DeepSeek | 0.818 | mean QA score (0-1) | static QA, mean over T1-T7 (secondary) | Reported by the authors | Paper Table 1 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 4 | DB-2.0-Lite | Not stated | 0.804 | mean QA score (0-1) | static QA, mean over T1-T7 (secondary) | Reported by the authors | Paper Table 1 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 5 | DB-2.0-Pro | Not stated | 0.787 | mean QA score (0-1) | static QA, mean over T1-T7 (secondary) | Reported by the authors | Paper Table 1 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 6 | Qwen3.6-Plus | Qwen | 0.783 | mean QA score (0-1) | static QA, mean over T1-T7 (secondary) | Reported by the authors | Paper Table 1 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 7 | GLM-5.1 | Zhipu AI | 0.757 | mean QA score (0-1) | static QA, mean over T1-T7 (secondary) | Reported by the authors | Paper Table 1 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 8 | Qwen3.6-35B-A3B | Qwen | 0.684 | mean QA score (0-1) | static QA, mean over T1-T7 (secondary) | Reported by the authors | Paper Table 1 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 9 | HY3-Preview | Tencent | 0.624 | mean QA score (0-1) | static QA, mean over T1-T7 (secondary) | Reported by the authors | Paper Table 1 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |
| 10 | Kimi K2.6 | Moonshot AI | 0.511 | mean QA score (0-1) | static QA, mean over T1-T7 (secondary) | Reported by the authors | Paper Table 1 | https://arxiv.org/html/2605.27887v6 | 9 Oct 2026 (checked 1 Nov 2026) |


Source: https://benchmarks.finance/benchmarks/portbench/

Created and maintained by Moneda.