AutoSignals - Autonomous Trading Signal Optimization

Monitors and controls the AutoSignals autonomous research loop.

What It Is

AutoSignals is an adaptation of Karpathy's autoresearch pattern for trading signal optimization. An autonomous loop runs continuously, spawning sub-agents to modify signals.py, backtesting changes, and keeping improvements.

Architecture:

• signals.py — The ONE file agents can modify (factor weights, thresholds, indicators, scoring)
• backtest.py — Fixed evaluation engine (5-year backtest, composite score metric)
• prepare.py — Data download (S&P 500 + held tickers)
• program.md — Instructions for research agents
• run.py — Autonomous loop controller
• experiments.jsonl — Full experiment log

Location: /Users/clawdiri/Projects/autosignals/

How to Use

Check Status

bash /Users/clawdiri/Projects/autosignals/status.sh

Shows:

• Running status (PID, uptime)
• Best composite score achieved
• Total experiments run
• Last 10 experiments with outcomes
• Score trend (last 20)
• Any errors

Start the Loop

bash /Users/clawdiri/Projects/autosignals/start.sh

Starts the autonomous loop in the background. Runs forever until stopped.

Stop the Loop

kill $(cat /Users/clawdiri/Projects/autosignals/autosignals.pid)

View Logs

tail -f /Users/clawdiri/Projects/autosignals/logs/autosignals.log

View Best Signals

cat /Users/clawdiri/Projects/autosignals/best_score.json

Then read the corresponding commit:

cd /Users/clawdiri/Projects/autosignals
git show <commit_hash>:signals.py

Monitoring Script (for DaVinci heartbeats)

bash /Users/clawdiri/Projects/autosignals/monitor.sh

Returns JSON with:

• running: bool
• experiment_count: int
• best_score: float
• best_commit: str
• trend: "improving" | "declining" | "flat"
• errors: list of recent errors

Evaluation Metric

composite_score = (0.35 * sharpe_normalized) + 
                  (0.25 * (1 - max_drawdown)) + 
                  (0.20 * win_rate) + 
                  (0.20 * profit_factor_normalized)

All components normalized to [0, 1].

Baseline targets:

• Sharpe: 1.57 / 1.46 / 1.24
• Starting weights: 40% Insider / 35% Earnings / 25% Sector Rotation

Good: Beat baseline Great: Sharpe > 2.0, drawdown < 15% Exceptional: Sharpe > 2.5, drawdown < 10%

Data

• Price data: 5 years daily OHLCV for S&P 500 + META, GOOG, AMZN, TSLA, BTC-USD, IAU
• Factor data: Currently mock (insider, earnings, sector). Can be enhanced with real API data.
• Cache: /Users/clawdiri/Projects/autosignals/data/prices.parquet

Refresh data:

cd /Users/clawdiri/Projects/autosignals
source .venv/bin/activate
python prepare.py

Design Principles (from Karpathy)