When restricted stock vests, many employees sell part of it straight away, often automatically, because too much of their wealth sits in one name. The company withholds shares to cover the tax and reports it on Form 4 with code F: a public, dated trace that a vesting happened. In a small company with a thin order book, that extra supply should push the price down for a few days. The sellers are not making a call on the stock; they are diversifying.
event-driven / insider flow
Employee vesting supply short
The idea
How it was tested
Three attempts. The first, on 2026-09-14, used recent small-float IPOs and produced only 25 trades with price data, under the floor of 30: void. The second, on 2026-09-15, used 300 small floats and came close: 137 trades, t +1.39 against 1.65, +1.600% per trade. Archived as a near miss.
The third was pre-registered on 2026-10-16 (commit 7f1e923) and declared the last. It replicated the hypothesis, not the sample: filings from 2014-01 to 2022-07 never examined for this idea, signals dated by the day the filing was published, and a universe of companies worth at most $400M on the event date itself. Short at the next open for 5 sessions, 2×ATR stop, 2.5×ATR target, fills re-priced when the open gaps past them. Two controls: a placebo on the same names 40 sessions earlier, and a calendar-time portfolio, because vestings cluster on the same dates.
Where it died
On 8,574 trades the alpha was −0.057% per trade, t −0.38. The calendar-time test gave t 1.01 and event minus placebo t 1.11, both short of 1.65. Gate 8 failed, and the halves disagreed (−2.25, then 1.24). Even a real effect would not have paid: the measured small-cap spread of 1.71% plus borrow meant each trade needed 2.11%.
The record
Thesis
When employees' stock vests, concentration risk pushes many of them to sell part of it right away, often automatically. The Form 4 code F filing (shares withheld by the company to pay the tax) is a public, dated proxy that vesting happened. In small names with little depth, the extra supply should weigh on the price for a few days.
Specification
- event
- Form 4, code F (shares withheld for tax, disposed), one event per (symbol, filing date)
- signal date
- first trading day on or after the filing's PUBLICATION date
- universe
- market cap on the event date ≤ $400M (historical, not today's)
- entry
- next open, short
- exit
- 5 sessions, 2×ATR stop, 2.5×ATR target, with fills re-priced when the open gaps past them
- window
- events filed 2014-01 to 2022-07, never looked at with this hypothesis
- controls
- Placebo — same names 40 sessions earlier with no event nearby; the event-minus-placebo daily series must have t ≥ 1.65. Also a calendar-time portfolio test, because vestings cluster on the same dates.
Gates
Pre-registrationpass
Was the hypothesis, its parameters and its pass/fail rule written down before any data was looked at?
Gate 7 — Jensen alphafail
After regressing out market exposure, is there alpha left?
t ≥ 1.65 per trade, in calendar time, and event minus placebo
t (HAC)-0.38alpha / trade %-0.06trades8,574t calendar time+1.01t event minus placebo+1.11
Gate 8 — Capture-preserved alphafail
Does the subset of trades a real portfolio can actually take keep the alpha?
t_cap ≥ 0.86 and ratio ≥ 0.50
t captured-1.84capture ratio+0.00
Gate F — Frictionfail
Is the alpha per trade large enough to pay for its own execution?
alpha/trade ≥ 0.30%
alpha / trade %-0.06
Operability was judged separately against a measured small-cap spread of 1.71% plus borrow: −0.06% per trade against 2.11% needed.
Stabilityfail
Is the alpha present in both halves of history?
t > 0 in both halves (2014-01 to 2018-06, 2018-07 to 2022-07)
t first half-2.25t second half+1.24
Why it died
With 8,574 trades dated by publication, code F filings do not anticipate a fall (t −0.38). The earlier near-miss of +1.39 does not replicate even in its own window once the universe is taken as of the event date and signals are dated by publication (t −0.44). Look-ahead does not explain it either: dated by transaction, as the original was, the effect is zero (t +0.03). The placebo did its job — shorting the same names without an event already gives beta 1.23 — and the event adds only a grey-zone t +1.11 on top.
What it taught
The near-miss came from a sample of 49 small floats picked by today's float (survivors) and signals dated by transaction, a median 2 days before anyone could read the filing. Replicating the hypothesis instead of the sample, on data never looked at, turned "almost significant" into nothing.
Verdict history
- 2026-09-14First pre-registration (recent small-float IPOs): only 25 trades with price data, below the floor of 30 — no verdict Void
- 2026-09-15Second pre-registration (300 small floats): n 137, t +1.39 vs 1.65, +1.600%/trade — archived as a near miss No alpha
- 2026-10-16Out-of-sample replication 2014-2022 with publication dating, point-in-time universe and placebo: t −0.38 No alpha