Does the fallback our studies use, the average with extreme returns clipped, tilt their readings?
Study 56 · Inconclusive · recorded 15 October 2026 · answer: Not settled · full paper with figures
The question
Since mid-October every study names in advance a fallback for a main measure too noisy to decide: the average after pulling the most extreme one percent of returns at each end in to the nearest kept value. It decided two published readings: the month after a late-report notice, and the extra recovery after a sharp one-day decline no filing explains.
Rules fixed before the result
Fixed in advance: where the clipped average sits is judged against a tolerance of a third of the effect each study named, one percentage point on the months and a third of a point on the stretches, and its error rate against a band of 2.5% to 7.5% around the 5% target. A range wholly beyond the tolerance reads as a shift, one across it as not settled; it needed 1,000 months, 150 companies and 10,000 stretches per price range.
Data and method
We rebuilt both earlier samples exactly: 7,777 ordinary months of 406 companies after an on-time report filing, each against comparable companies of the same industry and similar trading size, and 509,563 and 257,104 ordinary three-session stretches against the market. A return's upper tail reaches further than its lower, so clipping tends to lower the average; we measured how far, then drew 4,000 and 2,000 sets with no true difference copying each design and counted how often each range excluded zero.
The answer
Not settled. On the same companies' ordinary months the clipped average sat 0.45 percentage points below the plain average (95% range 1.25 below to 0.21 above), a range that crosses the one-point tolerance set in advance. Read with its registered range, it called a difference where none exists in 13.0% of tests (12.0% to 14.1%) against a target of 5%; the plain average did so in 6.95% (6.2% to 7.8%), and the clipped average tested against its own level in 8.35% (7.5% to 9.2%). On ordinary three-session stretches it sat 0.04 points lower (0.049 to 0.038) from prices of ten or more and 0.12 lower (0.132 to 0.107) from under ten, both within the tolerance of a third of a point, and its error rate was 5.45% (4.5% to 6.5%), within the band of 2.5% to 7.5%.
Main estimate
-0.005 fraction (95% range -0.013 to 0.002), from 7777 observations
What remains uncertain
These months and stretches were studied before, in studies fifty and fifty-five, whose recorded figures were seen when this plan was fixed. Conditions here are ordinary, with no event's heavier tails, and the drawn sets copy each design only approximately: with companies drawn with equal chance, the error rate rose for both averages, to about a fifth for the plain one and about one half for the clipped one. One development sample, 2014 to 2022.
How it was made
Run took 112 minutes; independent code review: approve; the recorded calculation reruns to the same result.
Fingerprints
plan d23d66814a4902796bee2ca7d19c9676a6218271287c305936f14dc2d78742c0 result 10e41adb17c73c640e882b80d783dc0fd3d6e848b5adecc79f6b0d72117ba151
Exploratory work on development data; not a performance claim.