Statistical arbitrage bets that deviations from statistical equilibrium relationships will revert. Unlike pure arbitrage (riskless profits from identical asset mispricing), stat arb bets on probabilistic relationships — there is genuine risk that the expected reversion does not occur. The 'arbitrage' is statistical rather than certain. A typical stat arb strategy might identify 500 cointegrated pairs, enter long/short positions when spreads exceed 2 standard deviations, and hold for days-to-weeks until reversion. Each individual pair contributes tiny, uncorrelated alpha; the portfolio-level Sharpe is built from diversifying across hundreds of independent bets.

The quantitative models underlying stat arb typically combine technical signals (relative momentum, mean reversion z-scores) with fundamental factors (relative valuation, earnings revision differentials) and risk factors (sector exposure, style factors) applied at high cross-sectional breadth. The fundamental theorem of active management (Grinold's law) states that Information Ratio = Information Coefficient × √Breadth — stat arb maximizes breadth (hundreds of bets simultaneously) to produce high Information Ratios from strategies with modest per-trade predictive accuracy (IC of 0.05-0.10).

Industrial stat arb requires significant technical infrastructure: real-time market data across thousands of securities, execution algorithms that minimize market impact when building hundreds of simultaneous positions, risk management systems that monitor exposure in real time and enforce concentration limits, and portfolio optimization engines that balance alpha signals against risk constraints in continuous time. The institutional barrier to stat arb is not the intellectual concept — it is the engineering and data infrastructure required to implement it at scale.

Execution quality is critical because stat arb alpha per trade is small. A stat arb strategy that generates 5 basis points per trade on average cannot afford to pay 3 basis points in bid-ask spread and market impact — the net alpha would be inadequate. Sophisticated execution algorithms (VWAP, implementation shortfall, dark pool access) minimize execution costs; direct market access and co-location reduce latency. Individual retail implementations of stat arb strategies face severe execution cost constraints that eliminate the thin alpha that institutional implementations capture efficiently.

The popularity of stat arb strategies creates crowding risk — when many funds hold similar positions based on similar models, simultaneous forced selling can create feedback loops. The 'quant quake' of August 2007 is the canonical example: multiple quantitative funds using similar factor models simultaneously faced redemptions and had to liquidate positions at the same time. Their models all had similar longs (quality stocks) and similar shorts (high-yield, low-quality names). When forced selling pushed these positions simultaneously, models designed to exploit mean reversion suffered losses that triggered further forced selling across the industry.

Signal decay is a continuous challenge: stat arb signals become less predictive as more capital exploits them, because the very act of trading on the signal partially corrects the mispricing. Strategies that worked with 100% of their historical information ratio 10 years ago may now work with 30-50% of that ratio as the signal has been competed toward equilibrium. Sustained outperformance in stat arb requires continuous research to discover new signals before they become crowded, retiring decayed signals, and innovating in data sources and model design.