English | 日本語

Qlib-with-Claudex

A quantitative investment R&D framework that autonomously drives Microsoft Qlib + RD-Agent using Claude Code + Codex.

Overview

Replaces OpenAI API dependencies with Claude Code subagents + Codex CLI, achieving a control inversion from "Python → LLM API" to "Claude Code → uses Python/Qlib as tools".

Repository Structure

Qlib/                          ← This repository (parent)
├── .claude/                   ← Claude Code config, skills, subagent definitions
│   ├── skills/                # RD loop, hypothesis generation, factor implementation, etc.
│   ├── subagents/             # Planner / Coder / Evaluator definitions
│   ├── settings.json          # Shared permissions
│   └── artifacts/             # Experiment artifacts
├── Qlib-with-Claudex/         ← microsoft/qlib fork (subproject)
├── RD-Agent-with-Claudex/     ← microsoft/RD-Agent fork (subproject)
└── docs/plans/                ← Design documents

Setup

# 1. Clone with submodules (single command)
git clone --recurse-submodules git@github.com:rozwer/Qlib-with-Claudex.git Qlib
cd Qlib

# If you already cloned without --recurse-submodules:
# git submodule update --init --recursive

# 3. Set up RD-Agent virtual environment
cd RD-Agent-with-Claudex
uv venv --python 3.12 && source .venv/bin/activate
uv pip install -e ".[dev]"

# 4. Install Qlib into the venv
uv pip install -e ../Qlib-with-Claudex/

# 5. Download Qlib market data (~50MB, CSI300 2005-2021)
cd ../Qlib-with-Claudex/scripts
python get_data.py qlib_data --name qlib_data_simple \
  --target_dir ~/.qlib/qlib_data/cn_data --region cn
cd ../..

# 6. Generate source_data.h5 (quick test)
python scripts/prepare_source_data.py --output /tmp/source_data.h5

# 7. Verify data quality
python scripts/check_data_quality.py /tmp/source_data.h5 /tmp/data_quality.json

Scripts

Script	Purpose
scripts/prepare_source_data.py	Generate source_data.h5 from Qlib market data
scripts/calc_ic.py	Calculate IC/IR/RankIC from backtest results
scripts/check_data_quality.py	Inspect column missing rates, output data_quality.json

# Generate source_data.h5 for a full R&D loop (5 rounds, 50 instruments, 2019-2020)
source RD-Agent-with-Claudex/.venv/bin/activate
python scripts/prepare_source_data.py --output_dir .claude/artifacts/rdloop/my_run --rounds 5

# Customize: 100 instruments, longer period
python scripts/prepare_source_data.py --output_dir .claude/artifacts/rdloop/my_run \
  --rounds 10 --n_instruments 100 --start_time 2015-01-01 --end_time 2020-12-31

R&D Loop

Claude Code sequentially invokes the following subagents to automate factor discovery:

Step	Component	Role
Hypothesis Generation	Planner (Agent tool)	TraceView analysis → propose new hypothesis
Code Generation	Codex CLI (codex exec --full-auto)	Generate factor calculation code
Execution	Bash (RD-Agent venv)	Run factor.py + calculate IC
Evaluation	Evaluator (Agent tool)	Analyze results → provide feedback

About Qlib

This project is built on top of Microsoft Qlib, an AI-oriented quantitative investment platform.

Qlib provides the full ML pipeline for quantitative investment: data processing, model training, backtesting, and covers the entire chain from alpha seeking to order execution. It supports diverse ML paradigms including supervised learning, market dynamics modeling, and reinforcement learning.

Key Features (from upstream Qlib)

• Quant Model Zoo — 20+ SOTA models (LightGBM, LSTM, Transformer, TRA, HIST, etc.)
• Dataset Zoo — Alpha158, Alpha360 for US and China markets
• Auto Workflow — qrun for end-to-end research workflow
• RL Framework — Reinforcement learning for order execution
• Online Serving — Deploy models with automatic rolling

Qlib Documentation

• Official Docs
• Tutorials (Notebooks)
• Original README

About RD-Agent

Microsoft RD-Agent is an LLM-based autonomous agent framework for industrial data-driven R&D. This project replaces its OpenAI backend with Claude Code subagents.

• RD-Agent Paper
• Original README

License

MIT (inherited from Microsoft Qlib / RD-Agent)