The post-earnings bounce did not survive our cost assumptions
What this does not establish That every post-earnings strategy has no edge. Under our sample and cost assumptions the net effect was economically close to zero; other execution conditions would need their own test.
How it is usually sold
As a swing trade of a few days after quarterly results: "buy the post-earnings dip" from one camp, "ride the post-earnings drift" from the other. Both come with a chart of one stock that did exactly that. Neither comes with the fifty thousand releases where it did not, nor with the transaction costs of a small account, which at this horizon eat more than the effect ever measured. The usual tools of the pitch are an earnings calendar with alerts, a gap scanner that lists the stocks that opened far from yesterday's close, and an oversold reading of an indicator such as the RSI on the day after the gap, presented as the buy signal.
Usually sold with Earnings calendar alertsGap scannersRSI oversold after the gap
Where we cast
Two opposite ideas about what happens in the ten days after a large US company reports quarterly earnings. The first is post-earnings drift, one of the most published anomalies in financial history: stocks that react very well to their results keep rising for weeks. The second is reversal: stocks that react very badly bounce back. Both were measured on 56,000 earnings releases since 1993, with the index membership rebuilt date by date so as not to look only at the companies that survived, with realistic transaction costs and zero tunable parameters: the window, the terciles and the horizon were fixed before anything ran.
Why it looked promising
Post-earnings drift has a classic explanation: investors react slowly to new information, and the price takes weeks to absorb a surprise. Reversal has the opposite explanation: the panic on announcement day overshoots and the price comes back. Both are plausible, both are documented, and that is why both were tested with a single correction for having looked twice.
What we found
Drift delivered an excess return of less than one basis point per event after costs. Under our sample and cost assumptions, that is economically zero. In large caps with current data we found nothing left to collect, which is what the literature on published anomalies predicts for the most famous one of all; whether anything remains under other execution conditions or in other universes is a separate question this test does not answer. Reversal delivered the opposite of what was expected: minus seven basis points per event. The tercile of worst reactions keeps falling, and in the last stretch of the sample it falls twice as hard. Buying the bounce after a bad quarter is stepping in front of the torpedo. And under the cost assumptions of a small account, the commission and spread per side modeled here, both variants lose between thirty and forty basis points per event. A different execution setup would need its own assessment.
The original minutes, in Spanish (secondary evidence)
| A-continuación (PEAD) | B-reversión | |
|---|---|---|
| Eventos | 12.993 | 13.263 |
| Exceso medio neto/evento | +0,9 bps | −7,2 bps |
| t_NW (mensual, lag 3) | 0,36 | −1,26 |
| Último tercio | +2,4 bps | −12,2 bps |
| Mono de fechas (máx-de-2) | p 0,86 | p 0,04 |
| Estrés 20 bps/lado | −29,1 bps | −37,2 bps |
| Gates (a,b,c,d) | 🔴🔴🟢🟢 | 🔴🟢🔴🟢 |
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
That information and trade are different things. The result says something real: the earnings loser keeps losing. But that information lives on the short side, which was not pre-registered, and it cannot be chased now because that would mean looking at the data and then inventing the rule. It stays on record as knowledge, not as a strategy. We also learned the toll of a small account: under the friction we modeled for little capital, only signals with a gross edge of forty or fifty basis points per event survive. Few of the effects in the published literature get there at that cost level.
Questions you may ask
Because the short side was not written down before the test. We saw the data and now we know where the effect lives; choosing that side today would be inventing the rule from the answer, which is how false positives are born. A short version would need its own pre-registered test, paying for every attempt this family already made, plus a borrow cost that at this size does not pay off. It stays on record as knowledge, not as a strategy.
It was real when it was measured, decades ago, across thousands of stocks. Once an anomaly is published, people trade it until it is gone, and the biggest, most watched companies are where it goes first. In large caps with current data, in our test, nothing measurable was left. The textbook is not wrong about the past; it is silent about today, and our result is silent about smaller stocks and other execution setups.
Yes, for the worse. With the friction we modeled for little capital, each event costs thirty to forty basis points before any edge shows up, so only signals with a gross edge of forty or fifty basis points per event survive. Few effects in the published literature get there at that cost level; a larger account with lower friction faces a different arithmetic. Picture paying a toll at every gate on a road where the prize behind each gate is smaller than the toll.
Check your reading
Three questions to tell what this entry established from what it did not.
Is post-earnings drift zero everywhere now?
We do not know. Under our sample of US large caps, our costs and our rule, the net effect was economically close to zero. Other universes and execution setups need their own test.
Which cost assumption drives the 30 to 40 basis points per event?
The commission and spread per side we modeled for a small account. A larger account with lower friction faces a different arithmetic.
Why not short the loser, since it keeps falling?
Because that side was not written down before the test; choosing it after seeing the data is how false positives are born.
What would bring us back
Nothing, in this form. The family of event and reversal trades in US stocks accumulated five attempts with different data and all of them died. The earnings calendar stays as a tool for something else: filtering when not to trade other strategies, not as a directional signal. A test of the short side would require a new pre-registration that pays for the family's accumulated multiplicity, plus a stock borrow cost that at this scale does not pay off.