Short-term momentum in large caps: no robust edge in our test
What this does not establish That momentum is absent in every market or specification. In the universe, period and implementation we tested, short-term momentum did not meet our criteria for a robust, tradable edge.
How it is usually sold
As "momentum works if you stick to liquid names": a monthly rotation in large caps, usually backed by a long-only backtest over the last fifteen years of US stocks. That backtest rides the market, not the selection. Once the portfolio is long-short and self-financed, the drift disappears and what is left is indistinguishable from picking at random. The usual tools of the pitch are relative strength rankings, top gainers lists sorted by last month's return, and momentum screeners with a volume filter that promise to keep only the "clean" trends.
Usually sold with Relative strength rankingsTop gainers listsVolume-filtered momentum screeners
Where we cast
A recent version of short-term momentum: buy last month's winners and sell last month's losers, but only among high-turnover stocks, the ones that trade the most volume relative to their shares outstanding. A 2022 paper argues that the old problem of one-month momentum, which actually reverses, disappears if you keep the high-turnover names, and that there the effect is positive even in the five hundred largest companies. It was replicated on the index rebuilt date by date, with turnover computed from shares outstanding as known at each point in time, seventeen years of monthly portfolios, costs charged on the actual turnover of each leg, and zero free parameters.
Why it looked promising
Stocks that many people trade absorb information faster and more smoothly; stocks nobody trades move in jumps and revert. If that is true, one-month momentum should work among the first group and fail among the second. The loser on the other side would be whoever sells high-turnover winners too early.
What we found
The long-short portfolio lost money net: minus eleven hundredths of a percentage point per month. But the revealing test is the shuffle: the signals were permuted two thousand times among the high-turnover stocks to see whether picking the winners differs from picking at random. The test could not tell them apart. About 37 percent of the shuffles matched or exceeded the gross result: the benchmark gave no evidence that the selection did anything a random pick could not. That is a statement about outcomes under the shuffle procedure, not the probability that the strategy is luck. This test did not provide convincing evidence that the ranking added value before costs; after costs, the tested portfolio lost money. And in the stretch after 2019, the only one the original paper could not see, the result is negative. Where the anti-decay criterion had to decide, it decided.
The original minutes, in Spanish (secondary evidence)
| Métrica | Valor | Gate | |
|---|---|---|---|
| L/S neto mensual | −0,113 % | >0, t_NW≥2 | 🔴 |
| t_NW | −0,34 | 🔴 | |
| Barajado CS (N=2.000) | p 0,37 (bruto +0,055 %) | <0,025 | 🔴 |
| Último tercio / 2019→ | +0,286 % / −0,169 % | ambos >0 | 🔴 |
| Meses | 204 | ≥120 | 🟢 |
Gate names, results and thresholds as written in the original test record. The English summary above is derived from the same rows. Units: pb = basis points per month; t_NW = t statistic with Newey-West standard errors.
What we learned
Two lessons. The first is about the null: each leg on its own gained a point and a half per month, and a test that only bought winners would have declared victory. That was market drift, which the self-financed portfolio cancels. A long-only backtest in US stocks from 2009 to 2026 validates almost anything; the right question is what remains once you subtract what the market gave away. The second is about data: the original effect lives in an academic database from 1963 to 2018 with an exact construction; with a shorter effective window and a different turnover measure, there is nothing. That a result does not replicate with other data is not the replicator's failure. It is information about how fragile the result was.
Questions you may ask
It was the market's win, not the selection's. In those years almost any basket of US large caps went up. A test that only buys winners cannot tell the two apart; a long-short portfolio can, because the market move cancels between the two baskets. What remained after cancelling was indistinguishable, in this test, from picking names at random. Imagine two runners on a moving walkway: both arrive early, and only the difference between them tells you who ran.
No. They measured it on an academic database from 1963 to 2018 with an exact construction. We measured it on the index rebuilt date by date, over seventeen years, with a different turnover measure and real costs, and in the stretch after 2019 that the paper could not see. That a result does not survive other data and other years is not the replicator's failure; it is information about how fragile the result was. A finding that needs one exact dataset to exist is not one you can trade.
We did not test them here. What we can say is that long-only momentum in US stocks was discarded earlier, and this test consumed the single condition that had been left open to reopen it. The whole family of momentum in stocks is closed for us: a new test would have to pay for all the accumulated attempts, not just its own. Other markets would be a different family with a different bar.
Check your reading
Three questions to tell what this entry established from what it did not.
Did we prove momentum does not exist?
No. In the universe, period and implementation we tested, short-term momentum did not meet our criteria. It says nothing about other markets or specifications.
What does the 0.37 from the shuffle mean?
That 37 percent of the random shuffles matched or beat our picks. The benchmark gave no evidence that the selection carried information; it is not the probability that the idea is luck.
Why did each side gain a point and a half per month and still count for nothing?
Because the market rose in those years. The long-short portfolio cancels that drift, and what remained was indistinguishable from random picking.
What would bring us back
Only genuinely new data. This test consumed the reopening condition that had been left open when long-only momentum was discarded. The whole family of momentum in stocks is now closed, and testing it again would require a pre-registration that pays for all accumulated attempts, not just its own.