The last article produced a rule for an honest stop. This one shows the day it cannot keep its word — and it fails worst at exactly the confidence level people trust most.
An event with a date on it
The stop here is built from the distribution of past returns. That distribution does not know what a calendar is: to it, tomorrow is one more day. But an earnings release is scheduled weeks ahead, everyone knows it is coming, and it is nothing like one more day.
96 NVDA earnings reactions, back to 2002. For each: build the threshold from the 250 sessions before the announcement, then check whether the reaction session breached it.
Reading the numbers
A stop labelled “one in a hundred” is taken out on one earnings day in eight. Counting both directions, 43.8% of reactions land outside the 95% band — against roughly 4% on ordinary sessions.
The uncomfortable part is the pattern: the error factor climbs from 6× at 0.90 to 26× at 0.99. A reader raises the confidence level to be safer, and that is precisely where it betrays them hardest.
How many sigma
Median 2.01σ. Half of all reactions are 2σ or larger, 13.5% are 5σ or larger. Under a normal assumption a 5σ event arrives about once in 3.5 million; here it arrives one time in seven. Not wrong by tens of percent — wrong by six orders of magnitude.
An earnings session is 4.5× as volatile as an ordinary one: 10.39% against 2.32%.
What no formula fixes
Look at the opening gaps of the last twelve: +6.58% · +11.19% · +7.45% · −3.38% · +5.52% · +5.06%.
The reaction arrives in the opening gap, not during the session. Which means a stop resting at X does not fill at X — it becomes a market order at whatever the open happens to be. On an earnings day a stop is not a stop; it is a prayer with a price written on it.
No refinement of the estimator repairs that. There are two answers: smaller size, or do not hold through the event.
This is the night a fund is absent and you are not
A fund going into an earnings release has usually cut the position, bought protection, or stepped aside entirely. It has a risk desk that requires this.
A private investor holds through — usually because they believe in the company, which is the correct reason to own a share. Nobody tells them that on that one night, the protective number on their screen stops meaning anything.
That is a real asymmetry, and it is not in the model. It is in who gets warned.
What this means for an instrument
An honest measurement has to know what it does not know. The past distribution contains no calendar, so the screen’s duty is to say so — rather than printing a confident-looking number for a day it cannot describe.