Reinforcement Learning Trading Bot

Multi-seed RL ensemble for buy/sell signals on tech equities (NVDA, AMD, MU) using Binary PPO with walk-forward validation.

Quick Start (5 minutes)

Windows (PowerShell):

py -m venv .venv
.venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt
streamlit run src/dashboard/main.py

macOS/Linux (Bash):

python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
streamlit run src/dashboard/main.py

Dashboard opens at http://localhost:8501

Current Status

Ticker	Status	Seeds	Sharpe	Alpha
NVDA	✅ Promoted	[3, 13, 7, 42]	2.03	+0.11–+0.52
AMD	✅ Promoted	[13]	2.01	+0.28
MU	✅ Promoted	[3, 7, 42]	1.77	+3.07

See CLAUDE.md for setup, architecture, and workflows.

Key Features

• Multi-seed ensemble — votes across 3+ models per ticker
• Walk-forward validation — train/val/test splits prevent look-ahead
• 6-gate promotion — actionable accuracy, win rate, drift, trade rate
• Exit manager — configurable position exit rules
• Streamlit dashboard — signal analytics, experiments, performance metrics
• Binary PPO — discrete 2-action space (Buy, Hold)

Common Commands

# Run dashboard
streamlit run src/dashboard/main.py

# Run experiment sweep
python src/experiments.py --ticker nvda --binary-actions --seeds 3,13 --timesteps 80000

# Evaluate sweep & promote
python scripts/evaluate_sweep.py --ticker NVDA --label nvda-exp10 --promote

# Test exit rules
python scripts/backtest_exit_rules.py --ticker amd --voting-method weighted

# Run tests
pytest tests/ -v

Documentation

• CLAUDE.md — Development guide, setup, workflows, architecture
• context-map.md — Full system architecture and data flow
• PROJECT_STATE.md — Current phase, promotion status, next steps