菜单
Use when an agent needs to review an existing factor library, summarize experiment logs, quantify acceptance rates and score dynamics, analyze factor families and correlations, and recommend the next research direction.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/XinAloha/skills --skill factor-review命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Use when an agent needs a standard cross-sectional long-only backtest protocol with T+1 open execution, top-bucket equal weighting, fees, limit-up or limit-down exclusions, benchmark comparison, NAV curves, drawdown, IC, and diagnostic charts.
Brinson-Fachler / BHB sector attribution with interaction, concentration (HHI), top contributors, and optional Carino multi-period linking. Use to explain active return as allocation vs selection.
Run and explain cross-sectional trading-factor backtests on schema-compatible factor signals and market data. Use when Codex needs to validate backtest inputs; calculate returns, NAV, drawdown, holdings, turnover, benchmark or hedged performance; generate a PDF diagnostic report; or explicitly verify the checked-in test fixture. Also use for Chinese requests such as 因子回测、收益验证、净值、回撤、持仓、换手、对冲收益或回测报告.
正在显示 SKILL.md
| name | factor-review |
| description | Use when an agent needs to review an existing factor library, summarize experiment logs, quantify acceptance rates and score dynamics, analyze factor families and correlations, and recommend the next research direction. |
| quantSkills | {"project_type":"skill","category":"tooling","tags":["factor-review","factor-library","research-review","correlation","workflow"],"platforms":["claude-code","codex","openclaw","cursor"],"status":"stable","validation_level":"listed","maintainer_type":"community","summary_zh":"不是单因子评价，而是因子库整体复盘 Skill：扫描实验日志 + 因子卡，输出三层报告（量化盘点 + 结构分析 + 研究建议），回答\"已经做了什么、最优在哪、下一步该挖什么\"。","summary_en":"Factor-library review skill for experiment logs, acceptance rates, score dynamics, factor-family structure, correlations, and research recommendations.","license":"GPL-3.0"} |
{
"version": 1,
"task": {
"placeholder": "请给出因子库/实验日志位置（如 runs.jsonl、因子卡目录）",
"required": true
},
"fields": [],
"prompt_template": "{{#task}}任务与材料：\n{{task}}\n\n{{/task}}{{#attachments}}用户上传的材料（已放入工作区）：\n{{attachments}}\n\n{{/attachments}}复盘任务材料中给出的因子库与实验日志，依次完成接受率与分数轨迹的量化盘点、因子族与相关性结构分析、下一步研究建议；每项建议写明 op_type、具体改动和预期分数，输出中文报告。"
}
跑了 N 轮迭代之后，到了"停下来想想"的时候。本 skill 把因子库 + 实验日志做一次系统性复盘，回答三个问题：已经做了什么、最优在哪、下一步该挖什么。
Layer 1: 量化盘点 — 跑了多少轮、接受率、分数轨迹（细则：references/quantitative-stats.md）
Layer 2: 结构分析 — 因子族分布、相关性矩阵、最优路径（细则：references/structural-analysis.md）
Layer 3: 研究建议 — 下一步该试什么、该停什么（细则：references/research-recommendations.md）
1. 扫描 runs.jsonl / INDEX.md / factor cards → 得到原始数据
2. Layer 1：算接受率、CRASH 率、分数轨迹、提分动力学
3. Layer 2：因子族归类、相关性矩阵、回溯最优路径的关键决策
4. Layer 3：基于 Layer 2 的发现给 3~5 个具体下一步假设
5. 输出标准报告（references/report-template.md）
| 概念 | 你的项目对应 |
|---|---|
| 实验日志 | journal/runs.jsonl / experiments.jsonl / git log |
| 因子库索引 | factor_library/INDEX.md / models/registry.json |
| 因子卡 | factor_library/<id>/card.json |
| best 标记 | journal/best.json / 当前 alpha.py |
最低要求：项目里至少要有"实验日志"（哪怕是 markdown 列表）才能复盘。否则先建机制再回头。
| 何时读 | 文件 |
|---|---|
| Layer 1 怎么算指标 | references/quantitative-stats.md |
| Layer 2 怎么做结构分析 | references/structural-analysis.md |
| Layer 3 建议格式 | references/research-recommendations.md |
| 标准报告模板 | references/report-template.md |
| 复盘反模式 | references/anti-patterns.md |
agents/openai.yamlagents/cursor-rule.mdcagents/portable-loader.md按 QUANTSKILLS 社区规则 §8 声明。