Skip to content

FinAI Research Workflow v1.0.0

Choose a tag to compare

View all tags
@csmar432 csmar432 released this 08 Jul 12:44
v1.0.0
f5d5b0c

FinAI Research Workflow v1.0.0

First Zenodo DOI release.

End-to-end AI research workflow for economics and finance — 8 stages from research idea to submission-ready manuscript, with human-in-the-loop checkpoints at every stage.

Highlights

  • 43 MCP data-source servers: A-shares (Tushare), US equities (yfinance), macro (FRED/World Bank/IMF/OECD), academic papers (OpenAlex/ArXiv/Semantic Scholar/NBER), SEC EDGAR, Eastmoney research reports, etc.
  • 47 econometric methods: Standard DID, modern staggered DID (Callaway-Sant'Anna, Sun-Abraham, Borusyak-Jaravel-Spinks), synthetic control, instrumental variables, panel GMM, RDD, mediation analysis, panel quantile, spatial regression, triple-diff DID, local projections, interactive fixed effects, robust wild-cluster bootstrap, Leamer bounds, Oster bounds, ...
  • 30 journal templates: 经济研究, 金融研究, 管理世界, JF, JFE, RFS, JPE, Econometrica, AER, JoE, etc.
  • 17 AI skills for Claude Code, Cursor, GitHub Copilot

What's tested

  • 8,048 pytest tests across 398 test files (32.2% coverage baseline)
  • All 47 econometric methods individually validated against analytical references (MAD ≤ 0.05)
  • 19-class robustness runner with JF/JFE/RFS-standard checks

Installation

pip install git+https://github.com/csmar432/finai-research.git
# Or:
git clone https://github.com/csmar432/finai-research
cd finai-research
pip install -e .

Quickstart

python scripts/start_research.py --topic "Carbon trading and green innovation"

Citation

See CITATION.cff. Zenodo DOI: to be assigned by Zenodo webhook within minutes of this release.

Bug fixes since 2026-07-08 audit

  • volatility_models.py::VolatilitySpillover.diebold_yilmaz: 4 silent bugs (statsmodels API drift, shape mismatch, pandas applymap deprecation, .loc slice) caused the main spillover path to silently fall back to correlation-based approximation. All fixed.
  • pyproject.toml pytest upper bound widened to <10.0 (was <9.0; depended on patched versions to satisfy lock).
  • Repository URL renamed from FinAI-Research-Workflow to finai-research (old URL still 301-redirects).

Limitations

  • All generated outputs (papers, reviews, design docs) require human review before submission. LLM-generated citations, statistical claims, and methodology descriptions must be verified independently.
  • Tested primarily on macOS Apple Silicon (M-series) and Ubuntu Linux. Windows has minor fallbacks (see event_monitor.py, sandbox.py).
  • LLM features require DeepSeek or compatible API key in .env.

Acknowledgments

  • 43 MCP servers built on top of open APIs (OpenAlex, FRED, World Bank, IMF DataMapper, OECD SDMX, Yahoo Finance, SEC EDGAR, Eastmoney, etc.).
  • Built with statsmodels, linearmodels, matplotlib, and pytest.
Assets 2
Loading