Post-earnings announcement drift is one of the oldest anomalies in the literature (Bernard and Thomas, 1989). After a large positive surprise, analysts anchored to their old targets revise slowly and passive money arrives late, so the price keeps rising for days or weeks. The counterparty is whoever sells to the slow revisers at yesterday's valuation.
event-driven / post-earnings drift
Post-earnings drift long
The idea
How it was tested
This comes from the lab's early phase, May 2026, before pre-registration and before alpha gates. Signal: an EPS surprise between +5% and +100%, 1 to 5 days after the report. Universe: US stocks at $10 or more with $5M of daily dollar volume. Buy at the next open, hold 5 sessions, with a 2×ATR stop and a 3×ATR target. At most 20 trades a day, filled in alphabetical order rather than ranked, a selection bias known at the time. The judgment was a Sharpe split: 2014-2019 as the design period, 2020-2025 held out, failing if the held-out Sharpe fell by more than 50%.
Where it died
The Sharpe went from 0.308 to 0.116, on 18,492 and 19,774 trades, a degradation of −62.3%. The win rate sat near a coin flip in both halves.
On 2026-05-27, once alpha against SPY had become a gate, the strategy was checked again retroactively, assuming a beta of 1. The result was not a weak positive but a clearly negative t of −5.24. It was never re-derived with Jensen alpha.
The record
Thesis
After a large positive earnings surprise, analysts anchored to old targets revise slowly and passive flows lag, so the price keeps drifting up for days to weeks.
Specification
- signal
- EPS surprise between +5% and +100%, 1 to 5 days after the report
- universe
- US stocks, price ≥ $10, dollar volume ≥ $5M
- entry
- next open
- exit
- 5 sessions, 2×ATR stop, 3×ATR target
- daily cap
- 20 trades/day, filled in alphabetical order (no ranking)
- caveat
- From the lab's early phase: not pre-registered, judged on an in-sample / held-out Sharpe split. The alphabetical daily cap was a known selection bias of that period.
Gates
Raw significance (legacy)fail
Are the strategy's raw returns statistically different from zero?
held-out Sharpe degradation better than −50%
sharpe 2014 2019+0.31sharpe 2020 2025+0.12degradation pct-62.30n 2014 201918,492n 2020 202519,774
A held-out Sharpe test, not a t-stat — the instrument of May 2026.
Market-adjusted alpha (legacy)fail
Does the strategy beat SPY over the same holding windows, assuming beta = 1?
alpha t ≥ 1.65
alpha t-5.24
Retroactive check, beta assumed = 1; not re-derived with Jensen alpha.
Why it died
Out-of-sample Sharpe 0.116 with a ~50% win rate in both halves: the signal did not separate winners from losers. The retroactive alpha check against SPY came out significantly negative (t −5.24).
What it taught
One of the best-documented anomalies in the literature did not show up in a quick, short-horizon implementation on this universe. The candidate explanations written at the time — 5 days is the weak tail of a 60-day drift, an upper cap that cut the large beats, an alphabetical daily cap — were never tested, because iterating on them would have been tuning after the fact.
Verdict history
- 2026-05-09Held-out test: Sharpe 0.308 in 2014-2019, 0.116 in 2020-2025 (−62.3% degradation) No alpha
- 2026-05-27Retroactive alpha-vs-SPY check: t −5.24 No alpha