When a stock's daily range, measured as Wilder's ATR in percent of price, jumps far above its own recent norm, something has shocked it. The bet is that the shock overshoots and part of the move is given back once it passes. Nobody is obliged to trade badly here; it is a statistical regularity, not a forced flow.
The idea did not come from reading a document. It came from the lab's factor delta scan, which compared how each factor's predictive power had changed over time. ATR% was the factor whose relationship had collapsed in 2026, so the pre-registration said in advance that it expected the test to fail.
How it was tested
Pre-registered on 2026-09-02, one of three candidates from the same scan, next to the volume spike reversal and the one-day return reversal. Signal: ATR(14)% above its 20-day rolling mean plus 2 rolling standard deviations. The run went short and held 5 sessions with a 2% trailing stop, although the pre-registration had described holding periods from 1 to 30 days. The decision rule was a Sharpe of at least 0.7 before the formal gates. One run.
Where it died
The single run lost money before any benchmark adjustment: Sharpe −0.38, t −1.86 over 5,911 trades. It was filed as no alpha, as predicted.
Then both of its foundations gave way. On 2026-09-30 the delta scanner that proposed it was found to ignore cross-sectional co-movement and was replaced. On 2026-10-30 the method audit read the backtest script itself: the rolling mean and deviation were computed over the whole factor panel at once, so each stock was compared with a number built from every other stock, not with its own history. The sibling one-day reversal test had already been caught and fixed for the same construction on 2026-09-10; this one was not.
A rule that never ran cannot be refuted, so the verdict moved from "no alpha" to void. The page keeps the numbers as a record of what the broken script produced.
The record
Thesis
Volatility spikes are temporary: when a stock's ATR as a percentage of price jumps far above its recent average, the move that caused it reverses. The pre-registration expected it to fail, because the lab's factor scan said the ATR% relationship had collapsed in 2026.
Prior: Expected to fail ("edge broken, likely permanent").
Wilder, J. W. (1978). New Concepts in Technical Trading Systems.
Specification
signal
ATR(14)% above its 20-day rolling mean + 2 rolling standard deviations
direction
short, as run (the pre-registration describes holding periods of 1-30 days; the run used 5)
exit
5 sessions, 2% trailing stop
caveat
The rolling statistics were written inside the signal's query string over the whole factor panel, not grouped by symbol — the same construction the sibling one-day-reversal test later found had not run as specified. The candidate itself came from the factor delta scanner that the lab later found to ignore cross-sectional co-movement. Treat the result as weak evidence.
Gates
Pre-registrationpass
Was the hypothesis, its parameters and its pass/fail rule written down before any data was looked at?
Raw significance (legacy)not evaluated
Are the strategy's raw returns statistically different from zero?
pre-registered decision rule: Sharpe ≥ 0.7, then Gates 1-8
t raw-1.86sharpe-0.38trades5,911
Raw per-trade returns, no benchmark adjustment — and not a valid measurement: the script computed the rolling statistics over the whole panel at once, not per symbol, so the rule tested was not the rule registered.
Why it died
Void, reclassified after the October 2026 method audit: the backtest script computed its rolling mean and deviation across all symbols together instead of per symbol, so it never ran the rule as registered. The numbers it produced are kept for the record only: Negative raw returns (Sharpe −0.38, t −1.86 over 5,911 trades) before any benchmark adjustment. Fading volatility spikes lost money in the episodes where volatility kept rising instead of reverting.
What it taught
The largest change found by a factor scan is more often a regime shift than an opportunity. This was one of three candidates from the 2026-09-02 scan, all rejected; the scan itself was later shown to produce absurd t-statistics, so the candidates were never well motivated in the first place.
Verdict history
2026-09-02Pre-registered from the factor delta scan; single run rejected No alpha
2026-09-30The delta scanner that proposed it is found to ignore cross-sectional co-movement and is replaced
2026-10-30Reclassified as void by the method audit: rolling statistics computed over the whole panel, not per symbol Void