quant-paper-agent

quant-paper-agent

A Claude Code plugin that turns a quantitative-finance arxiv paper into a production-ready event-driven strategy — in four composable skills that share a single papers/<arxiv-id>/ artifact tree.

  arxiv search   ->   PDF extract   ->   plan + backtest + benchmark   ->   event-driven strategy
       |                 |                         |                                |
  paper-search      paper-extract           paper-replicate                 strategy-translate

Each stage writes into papers/<arxiv-id>/, so you can enter the pipeline at any step — hand Claude a paper PDF you already have, or a vectorized backtest you already wrote, and pick up from there.

Install

/plugin marketplace add lucaswychan/quant-paper-agent
/plugin install quant-paper-agent@quant-paper-agent

Local development install (from a checkout):

/plugin marketplace add /path/to/quant-paper-agent
/plugin install quant-paper-agent@quant-paper-agent

Then install the Python dependencies in whatever environment you run Claude Code from:

pip install -r requirements.txt

Requires Python >= 3.9.

The four skills

Skill	Input	Output
paper-search	keyword / arxiv id / date range	papers/<id>/paper.pdf, meta.json
paper-extract	paper.pdf	note.md, metrics.json (reported metrics + core formulas)
paper-replicate	note.md, metrics.json	plan.md (halts for approval), code/{signal,portfolio,backtest}.py, nav.png, drawdown.png, replicated metrics.json, benchmark.md
strategy-translate	validated backtest + framework_spec.md	strategy/strategy.py + unit tests (placeholder until spec is filled in)

Quick examples

Once installed, invoke in natural language:

• "Use paper-search to find q-fin.PM cross-sectional momentum papers from the last 6 months, max 5."
• "Run paper-extract on 2403.12345."
• "Replicate the strategy in 2403.12345."
• "Translate the validated strategy for 2403.12345 to our event-driven framework."

Or call the bundled Python scripts directly — see each skill's SKILL.md for script-level invocation.

Hard rules the pipeline enforces

These come from the design post and are encoded in the skills:

1. paper-replicate Phase A halts for plan approval. A plan without confirmation is just a guess.
2. No parameter tuning to match the paper. Differences are recorded and explained, not sanded down.
3. No hidden trading costs / slippage unless the paper specifies them.
4. strategy-translate refuses to generate code until skills/strategy-translate/references/framework_spec.md describes real event types, strategy contract, order and position APIs, and clock semantics. It ships with a production-grade size_contracts() and unit tests for the notional-to-contracts + margin-check class of bug.

Troubleshooting

• PyMuPDF install fails on Apple Silicon — update pip first (pip install -U pip), or install from wheel: pip install pymupdf --only-binary=:all:.
• yfinance returns empty frames — yfinance's API drifts; upgrade with pip install -U yfinance and retry. If a ticker changed (e.g. BRK-B vs BRK.B), the skill will surface the error rather than silently producing an empty backtest.
• papers/<id>/metrics.json formulas look garbled — PyMuPDF's math font extraction is best-effort. paper-extract flags this and preserves raw page text for manual review; the downstream paper-replicate plan step depends on human judgment on formulas, not blind re-execution.
• strategy-translate refuses to run — expected until skills/strategy-translate/references/framework_spec.md is filled in with your team's event-driven framework contracts.

Layout

quant-paper-agent/
├── .claude-plugin/
│   ├── plugin.json
│   └── marketplace.json
├── skills/
│   ├── paper-search/
│   ├── paper-extract/
│   ├── paper-replicate/
│   └── strategy-translate/
├── requirements.txt
├── LICENSE
└── README.md

License

MIT. See LICENSE.