A fully autonomous statistical arbitrage system trading in Alpaca's paper environment. Currently deployed on AWS EC2 with $100K starting capital for a 90-day live forward test.
🟢 Paper Trading Active — The agent is live on Alpaca's paper account, executing trades autonomously during U.S. market hours. This is the forward-validation phase; performance is being evaluated on truly out-of-sample data. Results will be published at the end of the 90-day trial.
The system runs 24/7 across two nodes: a Research Node (MacBook Pro) that rediscovers opportunities after hours, and an Execution Node (Dockerized on AWS EC2) that trades them intraday via Alpaca's WebSocket.
Signal Generation — DBSCAN clusters correlated assets from a 110-ticker universe, then Johansen cointegration isolates mean-reverting spreads with statistically significant half-lives.
Trade Filtering — An XGBoost meta-labeler scores every candidate signal using fractionally differentiated features and microstructure dynamics. Only setups exceeding a dynamic probability threshold reach the order router.
Risk Control — Hierarchical Risk Parity (HRP) allocates capital across active spreads daily. A CUSUM filter on SPY monitors for regime breaks and blocks new entries during macro instability. Cooldown timers, EOD liquidation, and short-borrow checks prevent whipsaw and overnight gap risk.
In-sample backtest over 5 years (March 2021 – February 2026) using walk-forward lifecycle-aware cointegration discovery, share-based P&L accounting, realistic slippage and borrow costs, and Reg-T leverage (2.0x):
| Metric | Agent | SPY |
|---|---|---|
| Total Return | 31.58% | ~60% |
| Max Drawdown | -1.46% | -24.47% |
| RoMD | 21.64 | ~2.4 |
Parameters (Z=2.39, AI=0.56, PT=1.90, SL=1.75, Lev=2.0x) were selected via survival-constrained Monte Carlo optimization that rejects configurations with drawdown exceeding 20% of starting equity.
Caveats: This is an in-sample result — baskets were discovered on the same price history being tested. Real paper trading will likely show lower returns and higher drawdowns. The point of the 90-day trial is to measure the out-of-sample gap.
Training data sourced from Wharton Research Data Services (WRDS) — TAQ millisecond-resolution trades across all U.S. exchanges, January 2021 through February 2026.
Python · XGBoost · Numba · Statsmodels · Alpaca API · Docker · AWS EC2 · Tailscale · Git-based sync between nodes