Futures Statistical Arb

Regular trading typically predicts whether a single instrument will go up or down. Stat arb trades the relationship between two instruments, expecting temporary deviations to revert. It's market-neutral when done correctly - You can profit whether the overall market goes up or down.

Start with highly correlated, liquid pairs: ES/NQ (equity indices), GC/SI (metals), or ZB/ZN (treasuries). These have established relationships, good liquidity, and plenty of historical data.

You need enough to trade both legs with proper position sizing. For ES/NQ, minimum might be $50,000-100,000 to trade 1 contract of each with proper risk management. Micro contracts (MES/MNQ) allow starting smaller.

Yes. While stat arb is often lower risk than directional trading, losses occur when: The relationship breaks down permanently, The spread doesn't revert as expected, Execution slippage eats into profits. Risk management is still essential.

You need: Historical data for analysis (free from some brokers, or services like Quandl). Statistical software (Python with pandas/statsmodels, R, or Excel). Trading platform that supports spread orders or simultaneous execution.

The Engle-Granger test: 1) Regress Y on X to get hedge ratio. 2) Calculate residuals (spread). 3) Run ADF test on residuals. 4) If ADF p-value < 0.05, pair is cointegrated. Python's statsmodels has coint() function that does this automatically.

If correlation drops significantly (below 0.7), the relationship may be breaking down. Consider: Tightening your stop, Exiting the position, Not adding to the trade. Monitor rolling correlation and set alerts.

Roll positions before expiration - Typically 1-2 weeks before last trading day. Roll both legs to same expiration month (usually deferred month). Consider roll costs in your P&L. Some spreads (calendar) are based on expiration differences.

It depends. Dollar-neutral ensures equal exposure but may not reflect the statistical relationship. Hedge ratio from cointegration is more robust for mean reversion. Consider both and choose based on your analysis and goals.

Rolling parameters: Daily or weekly with 60-day lookback. Hedge ratio: Weekly or use Kalman filter for real-time. Cointegration test: Monthly or quarterly. More frequent updates capture changing relationships but can add noise.

Steps: 1) Screen for cointegrated pairs. 2) Select pairs with low correlation between spreads. 3) Allocate capital based on risk (equal risk or Kelly). 4) Monitor portfolio-level metrics. 5) Rebalance as relationships change. Target 5-10 diversified pairs.

Stat arb fails during: Fundamental regime changes (new relationship). Correlation breakdown (financial crisis). Crowded trades (too many doing same thing). Model decay (parameters become stale). Extreme events (relationships don't hold). Continuous monitoring and adaptation is essential.

ML applications: Pair selection (classify profitable pairs). Entry timing (predict optimal Z-score). Regime detection (identify favorable conditions). Start simple (Random Forest), validate extensively out-of-sample, and combine with fundamental reasoning.

Capacity is limited by: Spread liquidity (how much can you trade without impact). Number of viable pairs. Competition (more traders = less edge). Typically $10-100M for individual strategies before impact becomes significant. Larger capital requires more pairs or longer-term approaches.

During stress: Correlations often increase (all assets move together) then can suddenly break. Spreads can widen to extremes. Liquidity may dry up. Consider: Reducing position size, Widening stops, Avoiding new trades until volatility normalizes. Some stress creates opportunities, but risk is elevated.