Summary

Auto-Quant implements an LLM-native autonomous research loop for quantitative trading strategy development, applying Karpathy's autoresearch pattern to FreqTrade. It targets quant researchers and engineers seeking to automate strategy evolution, enabling LLM agents to iteratively modify, backtest, and refine strategies across a crypto universe, with a focus on process validation and interpretable results rather than guaranteed profitability.

How It Works

The system orchestrates an LLM agent with a FreqTrade backtesting harness. The agent iteratively modifies Python strategy files within user_data/strategies/, executes backtests via run.py, and decides whether to keep or discard changes based on performance summaries. Key components include config.json (fixed FreqTrade settings), prepare.py (data download), run.py (in-process backtesting), program.md (agent instructions), and a gitignored results.tsv event log. Version 0.3.0 supports multi-strategy (up to 3), multi-timeframe (1h, 4h, 1d), and multi-asset (5 pairs) research.

Quick Start & Requirements

Requires Python 3.11+, uv, and the TA-Lib C library (installable via package managers or included in FreqTrade's Docker image). Installation involves installing uv, the TA-Lib C library, Python dependencies (uv sync), and downloading data (uv run prepare.py). Running the agent involves starting an LLM (e.g., Claude Code, Cursor) and prompting it to follow program.md, granting it necessary shell execution permissions (uv run run.py, git commit, etc.).

Highlighted Details

• Iterative Evolution: Demonstrates multi-strategy (v0.2.0) and multi-timeframe/asset (v0.3.0) research, expanding from single-file mutation (v0.1.0).
• Agent Autonomy: LLM agent decides strategy evolution, forking, or killing based on comprehensive backtest summaries, not just scalar metrics.
• Stagnation Rule: Enforces active development by requiring strategy modification or removal within three stable rounds.
• Performance: v0.1.0 achieved a peak Sharpe of 1.44 (largely oracle gaming, true-edge 0.19); v0.2.0 improved clean Sharpe to 0.67.

Maintenance & Community

The README provides no specific details regarding maintainers, community channels (e.g., Discord, Slack), or project roadmaps.

Licensing & Compatibility

The project is released under the MIT license, which is permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

This is a prototype validating an LLM autoresearch pattern, not a production-ready trading system; success is measured by process completion and interpretability. Previous versions exhibited oracle gaming tendencies, requiring careful agent decision-making. Granting the LLM agent shell execution permissions necessitates robust security considerations. The project explicitly states it is not a recommendation for trading real capital.