QRAFTI: An Agentic Framework for Empirical Research in Quantitative Finance

Battle Count: 77

Directly relevant to quantitative trading as it automates core factor research workflows including signal construction, portfolio sorts, backtesting, risk adjustment, and robustness checks. The framework addresses the 'factor zoo' problem and replication crisis in empirical asset pricing. However, it is positioned as a research assistant tool rather than a trading system, and does not implement live trading, execution, or real-time portfolio management. The standardized diagnostics (alpha, t-stats, size quantile breakdowns) are directly applicable to strategy evaluation.

High complexity due to: multi-agent orchestration with three specialized agents, MCP server architecture for tool exposure, Panel data class with extensive operations (winsorization, quantile binning, rolling transforms, portfolio aggregation), reflection-based planning pipeline, context compaction mechanism, computation graph visualization, Pydantic-AI framework integration, OpenTelemetry monitoring, Streamlit UI, and CLI interface. Requires access to CRSP/Compustat databases and WRDS subscription. Uses Python 3.13 with specific dependencies.

Code, documentation, and Internet Appendix available on GitHub. Uses standardized CRSP/Compustat data, MCP-exposed tools for consistent computations, computation graphs for traceability, and structured logs. However, requires subscription-based databases (CRSP, Compustat via WRDS) and specific LLM (Gemini 2.5-Pro). Evaluation limited to two replication examples and one open-ended factor generation.