# KTD-Fin

End-to-end LLM trading on CSI300 A-shares, with tickers and dates masked so models cannot recall history. A Barra-style attribution splits returns into market, style and stock-selection alpha.

## Facts

- Maker: Stepfun; Tsinghua University (with FinStep, Shanghai Jiao Tong University, Adelaide University)
- Year: 2026
- Use case: Financial trading and portfolios
- Size: 10 frontier LLM agents plus 18 Qlib factor baselines; 548-trading-day window (2024-01-01 to 2026-04-10) plus 10 regime windows; 4 masking levels x 3 decision modes
- Public split: none found (paper says it is released as a template, but gives no URL)
- Grading: Total return and Barra-attributed selection alpha (leaderboard ranks by alpha); reliability panel includes ECE; Brier score is recorded for diagnostics only
- Licence: Not stated
- Submission: none
- Products allowed: Not stated
- Latest version: arXiv v1 (2026-05-27)
- What it does not cover: Markets outside Chinese A-shares, and inputs other than price data (observations are OHLCV only)
- Last checked: 1 Nov 2026
- Links: paper https://arxiv.org/abs/2605.28359
- Directory page: https://benchmarks.finance/benchmarks/ktd-fin/
- Leaderboard: https://benchmarks.finance/leaderboards/ktd-fin/

## Results

### KTD-Fin: total return % (higher is better, 10 models)

| Rank | Model | Provider | Score | Metric | Setting | Reported by | Source | URL | Date |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | Qwen3.6-Plus | Qwen | 85.3 | total return % | blinded x open-research, 2024-01-01 to 2026-04-10; alpha rank 2; CSI300 buy-and-hold +36.92 | Reported by the authors | Paper Table 4 | https://arxiv.org/html/2605.28359v1 | 27 May 2026 (checked 1 Nov 2026) |
| 2 | GPT-5.5 | OpenAI | 61.3 | total return % | blinded x open-research; alpha rank 6 | Reported by the authors | Paper Table 4 | https://arxiv.org/html/2605.28359v1 | 27 May 2026 (checked 1 Nov 2026) |
| 3 | doubao-seed-2.0 | ByteDance | 61.0 | total return % | blinded x open-research; alpha rank 3 | Reported by the authors | Paper Table 4 | https://arxiv.org/html/2605.28359v1 | 27 May 2026 (checked 1 Nov 2026) |
| 4 | Claude Opus 4.7 | Anthropic | 58.8 | total return % | blinded x open-research; alpha rank 1 | Reported by the authors | Paper Table 4 | https://arxiv.org/html/2605.28359v1 | 27 May 2026 (checked 1 Nov 2026) |
| 5 | minimax-m2.7 | MiniMax | 56.3 | total return % | blinded x open-research; alpha rank 4 | Reported by the authors | Paper Table 4 | https://arxiv.org/html/2605.28359v1 | 27 May 2026 (checked 1 Nov 2026) |
| 6 | step-3.5-flash | StepFun | 21.9 | total return % | blinded x open-research; alpha rank 5 | Reported by the authors | Paper Table 4 | https://arxiv.org/html/2605.28359v1 | 27 May 2026 (checked 1 Nov 2026) |
| 7 | Gemini 3.1 Pro | Google | 8.3 | total return % | blinded x open-research; alpha rank 7 | Reported by the authors | Paper Table 4 | https://arxiv.org/html/2605.28359v1 | 27 May 2026 (checked 1 Nov 2026) |
| 8 | deepseek-v4-pro | DeepSeek | 1.4 | total return % | blinded x open-research; alpha rank 8 | Reported by the authors | Paper Table 4 | https://arxiv.org/html/2605.28359v1 | 27 May 2026 (checked 1 Nov 2026) |
| 9 | GLM-5.1 | Zhipu AI | -3.5 | total return % | blinded x open-research; alpha rank 9 | Reported by the authors | Paper Table 4 | https://arxiv.org/html/2605.28359v1 | 27 May 2026 (checked 1 Nov 2026) |
| 10 | Kimi K2.6 | Moonshot AI | -24.2 | total return % | blinded x open-research; alpha rank 10 | Reported by the authors | Paper Table 4 | https://arxiv.org/html/2605.28359v1 | 27 May 2026 (checked 1 Nov 2026) |


Source: https://benchmarks.finance/benchmarks/ktd-fin/

Created and maintained by Moneda.