| name | build-b10-factor-evaluation |
| description | Use this skill to evaluate Alpha factors with IC/RankIC, ICIR, stratified backtests, turnover, decay curves, stock-pool filters, statistical diagnostics, and standalone HTML reports. |
| tags | ["quant","skill","build","factor-evaluation"] |
| license | GPL-3.0-only |
| maintainer | quantskills |
| runtime_adapters | ["codex","claude-code","cursor","hermes","openclaw"] |
| metadata | {"organization":"QuantSkills","organization_url":"https://github.com/quantskills","repository":"skill-build-b10-factor-evaluation","repository_url":"https://github.com/quantskills/skill-build-b10-factor-evaluation","project_type":"skill","collection":"build"} |
B10 Factor Evaluation Skill
Purpose
This repository packages the QuantSkills B10 factor evaluation BUILD as a reusable skill. It evaluates caller-provided Alpha factor panels and writes research and production artifacts without depending on a specific Alpha repository.
Use it when a user needs to:
- evaluate a factor's cross-sectional predictive power;
- compute IC, RankIC, ICIR, HAC t-statistics, decay curves, turnover, and stratified returns;
- compare quintile and decile grouping results;
- generate a standalone HTML factor-evaluation report;
- write production Parquet summaries for downstream agents.
Repository Layout
开发产物/SKILL.md: full development skill manual and Python API reference.
开发产物/skill.json: machine-readable metadata.
开发产物/scripts/: implementation, CLI runner, demo, visual report renderer, and tests.
开发产物/references/: API guide, changelog, and sample panels.
生产产物/SKILL.md: production artifact reader contract.
生产产物/数据库.parquet: production-format example/output database.
README.md and README.en.md: Chinese and English project introductions.
Inputs
The main input is a factor panel with these required fields:
| Field | Meaning |
|---|
trade_date | signal date |
ts_code | asset code |
factor_value | factor value |
forward_return | forward return aligned by the caller |
The implementation also accepts documented aliases such as date, asset, factor, and return. Multi-horizon forward-return columns may be supplied for decay analysis.
Usage
Run commands from 开发产物/ unless the caller has installed the scripts as a package.
python scripts/demo.py
python scripts/test.py
python scripts/daily_runner.py --input factor_panel.csv --output reports/ --production ../生产产物/
Python entrypoints:
from scripts.build import run, write_report, write_production
result = run(input_data, config={"group_mode": "quintile"})
write_report(input_data, "reports/factor_report.html", config={"target_id": "my_factor"})
write_production(input_data, "../生产产物/数据库.parquet", config={"target_id": "my_factor"})
Runtime Adapter Notes
- Codex and Claude Code can use this root
SKILL.md directly, then open 开发产物/SKILL.md for the complete implementation contract.
- Cursor should load
agents/cursor-rule.mdc.
- Hermes and other portable agents should load
agents/portable-loader.md.
- OpenClaw should load
agents/openai.yaml or agents/portable-loader.md.
Limitations
- This skill does not fetch market data or call PandaData directly. All raw market, industry, style, and forward-return data must be prepared before entering B10.
- It is an evaluation tool, not investment advice. Reports are research diagnostics and may not be presented as return promises or trading recommendations.
- Neutralization requires caller-supplied industry/style exposure panels.
- Production readers should read existing Parquet outputs and must not trigger recomputation during ordinary queries.
Compliance Metadata