Momentum's Decay Across Decades

A once-secret shortcut through a park stays fast only as long as few people know about it. Once it's on every map app, thousands of people take it at once, it gets crowded, and the time saved shrinks toward nothing. Momentum is the most famous version of this in finance: Jegadeesh and Titman published the 12-1 momentum effect in 1993, describing decades of data in which it had earned a large, robust premium. Every year since, more capital has known about that shortcut — and the premium earned by simply replicating their original rule has measurably shrunk.

Measuring the shrinkage

The standard test compares a signal's average return in-sample (the years the original study analyzed) against its average return out-of-sample (everything after publication), holding the exact construction rule fixed:

$\Delta = \bar{r}_{\text{post-publication}} - \bar{r}_{\text{in-sample}}$

In words: take the exact same 12-1 momentum portfolio construction — same lookback, same skip-month, same universe — and compare its average monthly return in the years after the paper came out to its average return in the years the paper studied. McLean and Pontiff (2016) ran this test across 97 published anomalies, momentum included, and found an average post-publication decline of roughly 26% relative to the original in-sample return, on top of a further, separate decline they attributed purely to statistical bias in the original studies (roughly another 10 percentage points), for a combined shrinkage close to a third of the originally reported effect.

Worked example 1. Jegadeesh and Titman's original 1965–1989 sample found the 12-1 winner-minus-loser portfolio earned about 1.0% per month, or roughly 12% annualized before costs. Extending the identical construction through the 1990s and 2000s, published replications have found the average monthly return closer to 0.6–0.7% — call it 0.65%, or about 8% annualized: a drop of roughly 35% in the raw monthly premium, consistent with the broader McLean-Pontiff finding, even before accounting for momentum's fatter tail risk (see Momentum Crashes) or higher realistic trading costs than the original paper assumed.

Two competing explanations

Worked example 2. Suppose a fund wants to know how much of the decline is "real decay" (arbitrage closing the gap) versus "estimation noise" (the original sample was unusually lucky). Take a simplified version: if the true underlying premium is 0.7% per month with monthly standard deviation 4%, a 25-year in-sample window (300 months) has a standard error on the mean of $4\%/\sqrt{300} \approx 0.23\%$. An original sample reporting 1.0% could easily be the true 0.7% plus about 1.3 standard errors of good luck — a statistically unremarkable draw, not proof of crowding. Distinguishing "the effect shrank because everyone started trading it" from "the original estimate was on the high side of a wide, noisy distribution" needs either a much longer post-publication window or an independent explanation for why the arbitrage mechanism (more capital chasing the same signal) should compress the premium — which McLean and Pontiff found direct evidence for: anomalies with better data availability and lower trading costs decayed more, exactly as a crowding story predicts.

What this means in practice

The practical response has not been to abandon momentum but to move past the textbook construction: residual momentum (see Residual Momentum), volatility-managed sizing, industry-neutralized variants, and shorter or longer lookback tweaks all exist partly because the plain-vanilla version that appeared in the original paper has a smaller edge today than it did in 1993. A researcher backtesting momentum today who reports the 1965–1989 Sharpe ratio as a forward expectation is quoting a number from a shortcut that has since become common knowledge.