How Statistical Arbitrage Decayed, 1990s to Today

Nunzio Tartaglia's quantitative group at Morgan Stanley in the mid-1980s built a strategy on a simple observation: stocks that move together most of the time occasionally drift apart for no fundamental reason, and when they do, betting on them coming back together makes money more often than not. Buy the stock that fell relative to its peers, short the one that rose, wait for the gap to close. It was called pairs trading, and later, generalized across hundreds of stocks at once, statistical arbitrage.

It worked spectacularly for years. It still works today. But its returns per dollar of capital, and even more sharply its returns per unit of risk, have fallen almost every decade since.

What actually decayed

Three forces did most of the damage. First, computing power: in the 1980s, finding correlated pairs and monitoring their spreads required real infrastructure that few firms had; by the 2000s, any desk with a laptop and historical data could replicate the basic idea. Second, decimalization in 2001 shrank the bid-ask spread stocks traded at from a sixteenth of a dollar to a penny, which directly reduced the round-trip cost of the same mean-reversion trades — good for execution, but it also meant the same dollar of gross alpha needed more capital deployed to matter, because per-trade edges shrank in tandem with spreads. Third, and most decisively, more funds ran the same family of strategies with similar risk models, so that when one large stat arb book needed to deleverage quickly, it sold the same stocks other stat arb books were long, and August 2007 became the strategy's defining crisis.

Worked example: August 2007

Many multi-strategy quant funds ran statistically similar equity market-neutral books — long a broad basket of "cheap relative to peers" stocks, short the "expensive relative to peers" ones. In the first week of August 2007, one large fund needed to raise cash unrelated to stat arb at all (its credit book was in trouble from the subprime shock). It sold its stat arb longs and covered its shorts fast. Because dozens of other funds held nearly the same positions, that forced selling pushed exactly the stocks a classic mean-reversion signal would have called "already cheap" even lower, and pushed the "already expensive" shorts even higher — the opposite of what the model expected. A fund with a normally stable book lost several percent in three trading days, then round-tripped most of it back within two weeks once the forced selling ended. The strategy hadn't stopped working; a shared crowded position had temporarily overwhelmed it.

What this means in practice

Modern stat arb desks survive by constantly refreshing signal universes (adding new instruments, alternative data, shorter or longer horizons), keeping leverage and factor exposures explicitly modeled so a common shock doesn't hit every fund's book the same way, and accepting that the fair return on a well-known, decades-old signal is now close to the cost of running it.