Autonomous trading strategy optimization on Bitcoin and commodities via HyperLiquid.

Inspired by Andrej Karpathy's autoresearch — an AI agent autonomously iterates on a trading strategy, backtests it, and keeps improvements.

An AI agent (Claude Code) runs in a loop:

1. Modifies strategy.py with an experimental idea
2. Runs uv run strategy.py to backtest on 90 days of out-of-sample data
3. If composite_score improves → keeps the commit
4. If worse → git resets and tries something else
5. Repeats indefinitely

All optimization runs in paper mode. No real funds are ever at risk during the loop.

# Install dependencies
uv sync

# Download market data from HyperLiquid (~5 min)
uv run prepare.py

# Run baseline strategy evaluation
uv run strategy.py

# Start autonomous optimization (in Claude Code)
# Just say: "read program.md and start the loop"

composite_score = sharpe_ratio x (1 - max_drawdown) x sign(total_return)

Rewards high risk-adjusted returns while penalizing drawdowns. Strategies with < 10 trades score 0.

• Source: HyperLiquid API (free, no API key needed)
• Assets: BTC, ETH (perpetual futures)
• Resolution: 1-hour candles
• Lookback: 365 days
• Split: first 275 days (train) / last 90 days (validation, out-of-sample)

# Run paper trading with the current best strategy
uv run execute.py

# Single evaluation
uv run execute.py --once

Live trading requires explicit confirmation and HL_PRIVATE_KEY environment variable.

Experiments are logged in results.tsv (untracked). Use analysis.ipynb to visualize progress.