Eaten by frictionArchived · long · n/a — never run
The idea
When a clearing firm fails to deliver shares it sold, federal rule makes it close the position before a fixed deadline, by borrowing or buying the shares, at whatever price the market asks. The obliged party did not write that obligation; the regulator did. Forced buying with a calendar attached should leave a trace in the price.
Of five forced-flow mechanisms examined on 2026-09-20, this was the only one that survived the entry test. The risk was declared up front: the retail-visible half of this ground, fails and "squeezes", is a heavily trodden story.
How it was tested
Rebuild the list of affected names from SEC fails-to-deliver data with the rule's own criteria: at least 10,000 shares failed and at least 0.5% of shares outstanding, for 5 consecutive settlement days. Operating companies only, price at least $1. Buy at the next open once a name becomes identifiable (day 5) and sell 8 sessions later, at the rule's day-13 deadline. Three predictions, all required: alpha t of at least 1.65, more alpha for larger fails relative to shares outstanding, and more alpha for longer-lasting fails.
Where it died
It never spent its single run. Before running it, the real spread was measured with live quotes on the 20 names of the universe: a median of 1.71%. If friction may take at most a third of the alpha, every trade had to earn 5.13%, or 0.642% a day over the 8 sessions the rule fixes. That is 8.8 times the best alpha per day this lab has ever demonstrated, 0.073%.
It had already been archived once, on a broken instrument: a Corwin-Schultz estimate put the spread at 0.77%. Calibrated the same day, the same estimator put AAPL, which trades at about one basis point, at 0.494%, and the verdict was retracted. The real spread turned out to be more than double the broken estimate.
The record
Thesis
A clearing participant carrying a fail-to-deliver must close it out by borrowing or buying shares before a deadline fixed by federal rule, whatever the price. That forced buying should leave positive abnormal returns in the affected name until the close-out deadline. The obligation is written by the regulator, not by the obliged party — 17 CFR 242.204(a): the participant "shall ... immediately close out its fail to deliver position by borrowing or purchasing securities of like kind and quantity".
Prior: The only one of five mechanisms examined on 2026-09-20 that survived the entry test. Declared risk: the retail-visible half of this ground (fails, "squeeze") is a heavily trodden narrative.
Specification
signal
Threshold-security condition reconstructed from SEC fails-to-deliver data with the rule's own criteria (fails ≥ 10,000 shares and ≥ 0.5% of shares outstanding, 5 consecutive settlement days), point-in-time share counts
universe
operating companies only (leveraged / inverse ETFs excluded), price ≥ $1, dollar volume ≥ $500K
entry
next open after the name becomes identifiable (day 5)
exit
8 sessions — from identification (day 5) to the rule's day-13 close-out deadline
predictions
P1 alpha t ≥ 1.65; P2 larger fails relative to shares outstanding give more alpha (controls for dilution); P3 longer-lasting fails give more alpha (controls for the borrow escape). All three required.
Gates
Pre-registrationnot evaluated
Was the hypothesis, its parameters and its pass/fail rule written down before any data was looked at?
Drafted 2026-09-20 but never locked — archived on a pre-run cost measurement. No run was spent.
Gate 7 — Jensen alphanot evaluated
After regressing out market exposure, is there alpha left?
t ≥ 1.65
Gate F — Frictionfail
Is the alpha per trade large enough to pay for its own execution?
round-trip spread ≤ 1/3 of alpha, over the rule-fixed 8-session holding, against the lab's demonstrated alpha ceiling of 0.073%/day
median spread pct+1.71required alpha per trade pct+5.13required alpha per day pct+0.64ceiling per day pct+0.07multiple of ceiling+8.80
Pre-run friction screen from live L1 quotes on the 20-name operating universe (20/20 valid), not the post-run 0.30% rule. Median 1.71-1.72% with or without one name carrying a stale, unadjusted price.
Why it died
The obligation is real; the economics are not. The median live bid-ask spread of the names on the threshold list is 1.71%. Allowing friction a third of the alpha means each trade would need 5.13%, or 0.642% per day over the 8 sessions the rule fixes — 8.8 times the best alpha per day this lab has ever demonstrated. Liquid names do not fail delivery, so the mechanism's own trigger selects exactly the spread that kills it. Stretching the holding until the arithmetic works would mean abandoning the rule's calendar and keeping only the ticker.
Open questions
The reconstruction of the threshold list never passed its own validation gate: entry precision / recall 0.733 / 0.698 against 0.80 required. Sixteen false positives remain unexplained after three named causes were ruled out. Moot for the verdict, left as measured.
What it taught
This one was first archived on a broken instrument. On 2026-09-20 a Corwin-Schultz spread estimate of 0.77% was used to archive it; calibrated the same day against names of known spread, the estimator put AAPL — which trades at about one basis point — at 0.494%, so the verdict was retracted. The real spread, measured with live quotes on 2026-09-30, was more than double the broken estimate: the instrument had understated the friction, not overstated it. A measurement that cannot tell AAPL from a $2 micro-cap cannot decide anything.
Verdict history
2026-09-20Passes the mechanism entry test; pre-registration drafted before the data pipeline existed
2026-09-20Validation of the reconstructed threshold list fails — entry precision 0.719, recall 0.651 vs 0.80; threshold not lowered
2026-09-20Archived on friction with a Corwin-Schultz spread of 0.77% Eaten by friction
2026-09-20Verdict retracted — the estimator puts AAPL at 0.494%, about 50 times its real spread
2026-09-29Reverse splits missing from the vendor's split data found and corrected — precision 0.733, recall 0.698, still below 0.80