The Avellaneda-Lee Stat Arb Framework

Instead of pairing up two individual stocks, the Avellaneda-Lee framework strips a stock's return down to the part that isn't explained by broad market risk factors, then trades the mean-reversion of that leftover, idiosyncratic piece. This scales pairs-trading logic up to hundreds of stocks at once using a shared set of systematic factors rather than one bespoke pair at a time.

How the signal is built

First, regress each stock's daily return on a set of risk factors, either statistically derived PCA factors from the whole universe's return covariance, or a simpler set of sector ETF returns, leaving a residual return series for each stock. That cumulative residual is modeled as mean-reverting, fitted to an Ornstein-Uhlenbeck process, which yields a standardized "s-score": how many standard deviations the current residual sits from its long-run mean, scaled by the estimated speed of mean reversion. Stocks with an s-score far below zero are flagged long candidates (unusually cheap relative to their factor-implied fair value); far above zero, short candidates.

Worked example

A stock's residual return, after removing PCA factor exposure, has drifted to an s-score of $-1.7$, its idiosyncratic price has fallen well below its estimated mean-reverting equilibrium relative to the market factors. The framework's trading rule (commonly entering when $s < -1.25$ and exiting near $s = -0.5$) flags this as a long entry, betting the idiosyncratic component reverts toward zero while the systematic factor exposure is hedged away by construction.