Avellaneda‑Lee revisited

A classic stock-picking rule still finds mean reversion in U.S. equities, but much of the edge disappears once realistic trading costs are added. (papers.ssrn.com) The framework comes from a July 11, 2008 paper by Marco Avellaneda and Jeong-Hyun Lee, which built market-neutral trades from stock-specific residuals — the part of returns left after common factors are stripped out. The paper tested two signal families, one based on principal component analysis and one based on sector exchange-traded funds. (papers.ssrn.com) In the original backtest, the principal-component version posted an average annual Sharpe ratio of 1.44 from 1997 to 2007 after transaction costs, while the exchange-traded-fund version posted 1.1 over the same span. Both weakened after 2002, and the principal-component strategy’s average Sharpe ratio dropped to 0.9 in 2003-2007. (math.nyu.edu) The basic idea is simple: if a stock moves too far from the pack after accounting for sector and market effects, the model bets that the gap will close. In practice, that means buying recent underperformers, shorting recent outperformers, and waiting for the spread to snap back. (math.nyu.edu) That setup has always had a catch: the signal tends to trigger lots of small trades, and small trades are where commissions, bid-ask spreads, slippage and shorting frictions pile up. Avellaneda and Lee wrote that holding periods in statistical arbitrage can run from seconds to days or weeks, which makes implementation costs central to the result. (math.nyu.edu) The original paper tried one fix by switching from calendar time to “trading time,” using volume to scale signals instead of the clock. In that version, the exchange-traded-fund strategy reached a Sharpe ratio of 1.51 from 2003 to 2007, better than the paper’s plain exchange-traded-fund signal. (papers.ssrn.com) The paper also tied the strategy’s swings to market structure, not just stock behavior. It said the performance around the summer of 2007 matched the “unwinding” explanation for the August 2007 quant drawdown described by Andrew Lo and Amir Khandani. (math.nyu.edu) That is why recent re-runs of the framework focus less on whether mean reversion exists and more on how quickly it resolves, how often a portfolio turns over, and how selective a trader has to be before costs erase the gross signal. The original paper’s own numbers already pointed in that direction: weaker post-2002 performance, better results when volume information was folded in, and sharp sensitivity during stressed liquidity periods. (papers.ssrn.com) The live-trading lesson is narrower than the old backtests made it look. A mean-reversion signal can still show up in the data, but whether it survives depends on the mechanics of execution as much as the math that found it. (math.nyu.edu)