My TradingAgent achieved 22.56% returns vs. DJI's 8.66%, a 2.318 Sharpe ratio, and 6.05% max drawdown over a 3-month live paper-trading evaluation.
My continual-learning schedule and outcomes were 2-hour fine-tuning cycles on a 48-hour sliding window (301-dim state, 8 technical indicators) with a 95% rollback threshold, achieving 52% acceptance across 271 cycles.
My TradingAgent interpretability and deployment approach: Integrated SHAP (global feature importance) and LIME (local per-trade explanations) for interpretable policy analysis; deployed end-to-end via the Alpaca API with turbulence-index-based position liquidation for live risk management.
I built TradingAgent, a PPO-based paper-trading project on GitHub. The README calls it production-ready, but describes paper trading and says it is for educational purposes only, not financial advice. My GitHub profile lists the repository on my github profile.
My TradingAgent repository is Jupyter Notebook, 1 star, 1 fork.
The system uses PPO; 30 US stocks; a 301-feature state; and 30 continuous actions in [-1, +1].
The state features are cash balance; 30 current stock prices; 30 stock holdings; and 240 technical-indicator features.
I use 8 indicators per stock: MACD, Bollinger upper and lower bands, 30-period RSI, 30-period CCI, 30-period DX, 30-period SMA, and 60-period SMA.
Trading runs in real-time at 1-minute intervals via the Alpaca API; orders are submitted using threading.
My stated risk controls are: liquidate positions when turbulence > 500; maximum 100 shares per stock; transaction costs 0.1% per trade; close all positions 15 minutes before market close.
Fine-tuning runs every 2 hours, using the last 48 hours of trading data and an 80% train / 20% validation split.
Fine-tuning uses 2,000 steps at learning rate 1e-5; I accept a model if new performance is at least 95% of original performance, otherwise reject it and roll back.
SHAP provides global feature importance; LIME provides local trade explanations; both are optional via --no-explain.
The README reports portfolio performance for Aug–Nov 2025: initial capital $1,000,000, final value $1,223,734, total return 22.5%, Sharpe ratio 2.318, peak portfolio value $1.26M, and maximum drawdown -6.05%.
The stated daily risk metrics are volatility 8.20%; best day +2.48%; worst day -2.43%.
Fine-tuning results were 271 total cycles; 141 accepted (52.0%); 130 rejected (48.0%); average improvement 1.53% per accepted fine-tune.
The three-month backtest without fine-tuning reports PPO +15.1% ($1.150M); TD3 +9.3% ($1.092M); DJI +8.6% ($1.086M); MVO +7.6% ($1.076M).
A separate README claim says PPO outperforms SAC by 4.5%, TD3 by 10.5%, and MVO by 10%, and beats DJI by 8.5%; this differs from the relative margins implied by the detailed comparison.
The stated periods are initial training: July 1, 2024–July 30, 2025; testing/evaluation: August 1, 2025–November 5, 2025.
The SHAP example gives AAPL_macd: 0.34; MSFT_rsi_30: 0.28; JPM_cci_30: -0.21.
The example LIME trade is BUY 45 AAPL @ $175.23, with contributions from AAPL_macd (+0.15), AAPL_close_30_sma (+0.12), and market_turbulence (-0.05).
The README says disabling explainability with --no-explain is 15% faster.
Repository metadata says Started 2026-01, last updated 2026-06.
The repository topics are deep-reinforcement-learning, drl, drl-trading-agents, explainable-ai, fintech, lime, ppo, reinforcement-learning, sac, shap, stablebaselines3, td3, trading-automation, trading-signals, and xai.
The README describes the project as “a production-ready paper trading system” and says it is for paper trading and educational purposes only, not financial advice.
The README says the dashboard's cumulative returns peaked at 27% and closed at 22.5%.
The README describes daily returns as a normal distribution centered at 0.
The README says the system maintained ~$1.2M in positions.
The README describes a linear correlation between original and fine-tuned scores.
The README says the rolling acceptance average was around 52% over time.
The README says accepted models correlate with portfolio growth periods.
The README says rejected models protect against degradation.
The default PPO model path is trained_models/agent_ppo.zip.
The default output directory is production_paper_trading_results.
The default trading interval is 1 minute.
The default fine-tuning interval is 2 hours.
The project requires Python 3.8+ and an Alpaca Paper Trading Account (free).
The README estimates ~500MB RAM without explainability and ~800MB RAM with SHAP/LIME.
The listed dependencies include stable-baselines3 for PPO, finrl for financial reinforcement learning, alpaca-trade-api for market data and execution, shap, lime, pandas, and numpy.
The project license is MIT; the README directs readers to the LICENSE file for details.
The README warns: “Use at your own risk. Past performance does not guarantee future results.”
The performance period predates the repository start date and the sample CSV timestamps, which are in January 2026; the README does not explain this chronology. It also gives both 22.56% and 22.5% return figures without explaining the precision difference. The claimed outperformance percentages do not match the detailed three-month comparison, and no SAC comparison table is provided. The README does not establish live-trading deployment or real-money results.