Fails-to-deliver, the threshold list, and the short-squeeze folklore

The folklore runs like this. A stock shows up on the Reg SHO threshold list, which means its fails-to-deliver have been large and persistent. Retail lore reads that as naked shorting: shares sold that were never delivered, which must eventually be bought back. So short interest has to unwind. We turned that into a claim we can falsify: after a symbol lands on the threshold list, its reported short interest does not fall more over the next two settlement periods than it does for comparable symbols that are not listed.

The data rejects that claim, but not in the way the folklore would like. Listed symbols' short interest does fall more, by a median of 12.6 percentage points against matched controls. Almost all of the gap comes from one group of listings, the ones where short interest had more than doubled in the period before. Everywhere else the listed and unlisted symbols look the same.

Universe and method

Three feeds. The threshold list has 91,525 rows from 2020-01-02 to 2026-10-23, and only 33,303 of them carry a `symbol_id` (1,222 symbols, against 3,589 distinct tickers on the list). FINRA short interest has 3,196,978 rows from 2020-01-15 to 2026-09-30, of which 1,023,344 are mapped to a `symbol_id`. SEC fails-to-deliver has 5,129,322 rows from 2016-01-04 to 2026-09-30, all mapped. We join only on `symbol_id`, never on ticker, so every unmapped row is out of the analysis. That is a large cut and the first caveat below.

An entry event is the first day a symbol appears on the list after at least 10 trading days off it. We drop spells that begin in the first 10 days of the data, where we cannot see the absence. That leaves 2,507 listing spells. Short interest is published for settlement dates about twice a month, so the baseline t0 is the last settlement date on or before the entry day (within 16 days), and the outcome is the change in `short_interest` from t0 to t+2, two settlement dates later. We require a short-interest row at t0, t+1 and t+2, with consecutive settlement dates. That leaves 2,246 events across 1,094 symbols, with entry dates from 2020-01-16 to 2026-09-14.

Controls are symbol-periods with the same t0 and no listing day from 21 days before t0 through t+2: 909,779 of them. For each event we take the median change among controls at the same t0 in the same days-to-cover decile and the same decile of fails-to-deliver dollars over the prior 28 days (the sum of `fail_quantity * price` over settlement dates up to t0; symbols with no fails get their own bin). Cells with fewer than 10 controls are dropped, and 2,235 events survive. The price is used only to size fails in dollars. We make no claim about prices or returns.

The headline

Median change in short interest from t0 to t+2, with the interquartile range:

• Listed symbols (2,235 matched events): median -9.8%, IQR -63.1% to 50.8%.
• Matched controls (the control median for each event's cell): median 2.0%, IQR -2.5% to 9.2%.
• Event minus its own matched control: median -12.6 points, IQR -65.9 to 47.7.

56.7% of listed symbols reported lower short interest at t+2, against 37.8% of their matched controls, and 58.1% of events came in below their control (the standard error on that share is 1.1 points). Without any matching, the picture is the same: 2,246 events at -10.1% (IQR -63.2% to 50.6%) against 909,779 controls at 0.0% (IQR -15.2% to 19.3%). The listed distribution is far wider than the control distribution, which is mostly cells of ordinary stocks.

Matching on days-to-cover and fails would be enough if listed symbols were otherwise ordinary. They are not. Going into the listing, the median listed symbol had short interest up 27.4% over the prior period, against 0.0% for controls, and 68.9% of events sat at the days-to-cover floor of 1.0 (FINRA does not report below it) against 28.6% of controls. Symbols whose short interest has just jumped tend to give it back whether or not they are listed, so this needs its own control.

The run-up matters more than the list

We re-matched on t0, the days-to-cover decile and the size of the prior-period change in short interest, in five classes: down more than 20%, down 0 to 20%, up 0 to 20%, up 20% to 100%, and more than doubled. Fails dollars drop out of this match. 2,145 events have a prior-period figure and a cell with enough controls.

• Down more than 20% (402 events): listed +38.0% against controls +33.2%, a difference of 8.5 points.
• Down 0 to 20% (283): 1.9% against 2.0%, a difference of -0.6.
• Up 0 to 20% (334): 1.3% against 0.0%, a difference of -0.4.
• Up 20% to 100% (386): -3.9% against -2.3%, a difference of -2.6.
• More than doubled (740): -60.8% against -25.8%, a difference of -34.9 points, with 69.5% of events below their control.

Across the first four rows, 47.5% to 51.8% of events fall below their control. That is a coin flip. The effect is in the last row. Short interest that has more than doubled reverts hard even without a listing: among the 59,031 unlisted symbol-periods in that class, the median change over the next two periods is -24.5%. Listed symbols in the class give back 60.8%, and that gap of 34.9 points is the part of the folklore that survives. Pooled over all classes the difference is -8.7 points (-9.7% against -1.5%).

A second match, on days-to-cover decile, fails dollars and a tercile of the prior change together, gives the same answer for the pooled sample: 2,059 matched events at -11.5% against -1.0%, with a median difference of -12.3 points and 57.6% below control.

Fails size cuts the other way at the low end

Splitting the first match by the size of fails over the prior 28 days:

• No priced fails (19 events): listed +125.2% against controls 0.0%.
• Lowest three deciles (205): +73.6% against +4.6%, a difference of 72.2 points.
• Middle four deciles (620): +11.7% against +2.2%, a difference of 8.5.
• Top three deciles (1,391): -20.0% against +1.8%, a difference of -22.7.

Listings with small recorded fails are followed by short interest going up, not down. Only the large-fails listings unwind, and they are the bulk of the sample. So "on the list" is not one thing. A big fail balance with a short-interest spike behind it is where reported shorts fall back; a listing with modest fails is not followed by covering at all.

Days-to-cover tells the same story where it can be measured. For the 665 events that start above the 1.0 floor, the median change over the two periods is -16.7% (IQR -45.1% to 35.4%) against 2.6% for matched controls, a difference of -14.8 points. In the first period after the listing, t+1, the gap is small: -2.8% against -1.6%, a median difference of -0.5 points. The unwind shows up in the second period.

How long listings last

Among 2,466 spells that began after the warm-up and ended before the last published day, the median listing is 10 trading days (IQR 6 to 16, 90th percentile 26, maximum 105). 18.9% last 5 days or fewer, 44.9% last 6 to 12, 25.4% last 13 to 25 and 10.8% last longer than 25. A further 41 spells were still open on 2026-10-23 and are excluded. Thirteen consecutive settlement days on the list triggers the close-out requirement for the fail, which is why a day count near 13 is a natural boundary. Most listings end before it.

ETFs versus single stocks

`symbols.asset_type` is empty for almost every row, so we classify by security name: names with words like ETF, Fund, Trust, Shares, Index, Leveraged or a sponsor brand count as ETF-like. It is a heuristic and it will mislabel some trusts and funds-of-funds. On that classification ETF-like securities hold 25.5% of fails-to-deliver dollars since 2020-01-02 (654 symbols, 0.744 trillion dollar-days of fails against 2.176 trillion for the rest), but only 7.1% of listing entries (179 of 2,507). ETFs carry a quarter of the fails dollars but are rarely the ones that stay failed long enough to be listed. We did not investigate why. Of the 9,443 symbols with fails in that window, 1,220 ever enter the list, and they account for 24.0% of fails dollars.

Querying it

The short-interest and fails-to-deliver feeds are public endpoints. The first call below pages short-interest history for one ticker; the second ranks a settlement date by the size of open fails, which is the screen that precedes a listing:

# Reported short interest, two latest settlement periods
curl -H "Authorization: Bearer $HEDGEFRIEND_KEY" \
  "https://api.hedgefriend.dev/v1/market/short-interest/GME?limit=2"

# Largest open fails on the most recent published settlement date
curl -H "Authorization: Bearer $HEDGEFRIEND_KEY" \
  "https://api.hedgefriend.dev/v1/market/fails-to-deliver?limit=20"

# One ticker's fails history
curl -H "Authorization: Bearer $HEDGEFRIEND_KEY" \
  "https://api.hedgefriend.dev/v1/market/fails-to-deliver/GME?from=2026-01-01"

{
  "data": [
    {
      "settlement_date": "2026-09-30",
      "short_interest": 42673721,
      "previous_short_interest": 39069126,
      "avg_daily_share_volume": 10727887,
      "days_to_cover": "3.98"
    }
  ]
}

The fails records carry `fail_quantity`, `price` and `fail_value_usd`, the notional used in the matching above. The Reg SHO threshold list is in our database but has no public route yet, so the entry events here were computed from the table, not through the API. The queries are in the research folder for this post.

Caveats

• Symbol mapping. Only 33,303 of 91,525 threshold-list rows and 1,023,344 of 3,196,978 short-interest rows carry a `symbol_id`, and joins on ticker are excluded by design. If the unmapped listings differ from the mapped ones, for example by being delisted or foreign-listed securities, the result does not describe them.
• Survivorship. An event needs short-interest rows at t0, t+1 and t+2. A symbol that is delisted, acquired or halted before t+2 drops out, and those may be the ones where short interest ended for a reason other than covering.
• Timing. Short interest is a snapshot on a settlement date, published after it. The entry day falls anywhere up to 16 days after t0, so the first period after entry is partly before the listing took effect. t+2 is two snapshots after the baseline, not a fixed number of days after the entry.
• Fails are open balances. A `fail_quantity` that stays elevated is one position persisting, and the published file is semi-monthly, so the 28-day fails sum is approximate. 23,179 fails rows have no price and are excluded from the dollar sums.
• Days-to-cover is floored at 1.0, which is why 68.9% of events show no measurable change in it; we report it only for events above the floor.
• No causal reading. Short interest falling after a listing is consistent with covering, but it is also consistent with share-count changes, restructurings or a spike that was already reversing. We cannot tell them apart and we do not serve prices, so we do not try.

What it supports

The folklore half-holds. A listed symbol's reported short interest does tend to fall over the next two settlement periods, by more than a matched unlisted symbol's, and it is not explained by days-to-cover, the size of the fails or a coarse prior-change match. But the effect is concentrated in listings that follow a doubling of short interest, a group that reverts by about a quarter even without a listing. For listings that follow a flat or modest change in short interest, the data shows no unwind. It does not support a rule of the form "threshold list entry means shorts must cover". It supports a narrower one: a threshold entry on top of a short-interest spike and a large fail balance is usually followed by that short interest falling back.