The 50% Short Volume Myth: 11.9 Million Rows of the Daily Short-Sale File, and What the Ratio Actually Predicts

TL;DR - Every retail short-squeeze screener publishes a "short volume ratio": the day's short-marked volume divided by total volume, lifted from a file FINRA posts each evening. The lore says above 50% means a stock is heavily shorted, that a rising ratio on a falling price is "them shorting it down", and that both set up a squeeze. We tested all of it on every row of that file from 3 January 2022 to 14 October 2026 - 11,867,351 stock-days across 1,200 sessions - joined to the consolidated tape, with the hypotheses hashed before any return was computed. In stocks trading more than $100M a day, the ratio is highest on up days (top decile +20 bp on the day, bottom decile -21 bp) and the top decile then earns 7 bp over the next five sessions against 35 bp for the bottom: -26.7 bp, t = -4.1, negative in both halves of the sample and in all five calendar years. The only tier where the lore's sign appears is stocks trading under $1M a day, where the gap is smaller than the quoted spread, two-thirds of it is the next morning's rebound from a closing print, and it has faded to t < 1.4 since 2025. The ratio itself is a coin that lands near 50% for everyone: above 50% on 52% of all stock-days, median 50.9%, crossing the line from one day to the next 29% of the time, with 83% of its variance explained by neither the stock nor the day. And the file it comes from covers only the off-exchange tape - 32% of a liquid stock's volume, median. Three things survive: short-marked volume is liquidity provision into buying, not a bet; a high reading is mildly bearish for the week in liquid names, not bullish; and nothing here is tradable net of spread in either direction.

Open any retail short-squeeze screener and the first column is "Short Volume %". GameStop 61%. Some biotech 72%. The colour scale turns red above 50%, and the caption, give or take, reads: the percentage of today's volume that was short selling. Half a dozen paid services sell it as a daily short-interest proxy, and the chat rooms treat a reading above 50% as proof that the stock is under attack.

The number is real. It comes from a file FINRA publishes every evening under Regulation SHO, one row per symbol: short volume, short-exempt volume, total volume. It is free, it goes back to 2009, and anyone can download it. Which also means anyone can test what it predicts. So we did, on all of it.

The set-up

We took the FINRA Consolidated NMS daily short-sale-volume file for every session from 3 January 2022 to 14 October 2026: 1,200 files, 11,867,351 rows, 18,901 symbols. Each row was joined to that symbol's daily consolidated bar for the same session, and to a security master built from the listing directory, the broker asset master and the SEC's registrant list, so that ETFs, warrants, rights, units, preferreds and notes could be dropped. That leaves 7,505 common stocks and ADSs. With a prior close of at least $1 and a bar on the tape, the panel is 5,722,364 stock-days in 7,448 names.

Three definitions carry the article:

• The ratio. Short volume divided by total volume, exactly as the file gives it and exactly as the screeners publish it. We call it the short-volume ratio.
• The sort. Every session, every stock in the universe is ranked on that day's ratio and cut into ten equal groups. The return of each decile is the equal-weight mean of its members. The headline statistic is the top decile minus the bottom decile, one number per session, 1,200 sessions, with Newey-West standard errors on that daily series.
• The tiers. Each stock-day is tagged with the stock's trailing 20-session median dollar volume, cut at $1M, $10M and $100M. The cuts were fixed before any return was tabulated. Below $1M a day is where the lore lives. Above $100M a day is where anyone with real size actually trades.

The hypotheses, the sort, the tiers, the inference and the bar for calling anything tradable were written down and hashed before the first return was joined. Two of the five registered predictions were wrong, and they are reported below in full.

Here is what the ratio looks like in the stock it is most often quoted for.

What the number is, before what it predicts

Start with the distribution, because the lore's threshold does not survive it.

In stocks trading at least $1M a day - 3,757,812 stock-days, 6,020 names - the median daily ratio is 50.9%. It is above 50% on 52.1% of all stock-days, above 60% on 28.5%, above 70% on 11.3%. The tenth percentile is 29.8% and the ninetieth is 71.1%. "Above 50%" is not a flag. It is the median stock on the median day.

It is also not a property of the stock. Take the 4,269 liquid names with at least 120 days of history and compute each one's own median: 53.2% of them sit above 50%, and the middle 90% of stocks have medians between 38.8% and 61.6%. The spread across stocks is narrow. The spread within a stock from one day to the next is not: the median name's day-to-day standard deviation is 14.4 points, and on 29% of days a stock is on the other side of the 50% line from where it was the day before. GameStop, above, crosses it on 22% of days.

We registered the opposite prediction: that the ratio would be a fingerprint, with most of its variance belonging to the stock, because a name's mix of wholesaler flow and institutional dark-pool flow ought to be stable. It is not. A two-way decomposition over the liquid panel gives the stock 15.3% of the variance, the calendar day 2.0%, and leaves 82.9% in the residual. The autocorrelation is 0.57 at one day, 0.29 at twenty, 0.22 at sixty and 0.14 at a year. A stock's average ratio in the first half of the sample explains less than a quarter of its average in the second half (correlation 0.48). Whatever the ratio carries, it is mostly noise around one half, and the one half is structural.

Why one half? Because of what a short mark is. Every trade has a seller, and the seller is marked short if it does not hold the shares. The firms that sell most often without holding shares are the ones whose job is to sell on demand: the wholesalers that fill retail buy orders from inventory they do not have, and the market makers posting offers in dark pools. The short mark is not a bet that the stock goes down. It is the footprint of someone providing liquidity to a buyer. A day with more buyers than usual gets more short-marked prints, not fewer. Which is the first thing the data says.

Finding 1: short volume is highest on up days

The lore's second claim is directional: a high ratio on a down day is "them shorting it down". Sort the >$100M-a-day names on the ratio and look at the same day's return.

Decile of the day's ratio	1 (lowest)	2	3	4	5	6	7	8	9	10 (highest)
Same-day return, mean, bp	-20.9	-9.0	-3.6	+0.3	+2.8	+6.1	+10.5	+13.8	+17.1	+20.4
Same-day return, median, bp	-12.1	-8.8	-5.3	-0.4	+1.8	+6.1	+9.5	+14.2	+17.5	+20.3

Monotone across all ten deciles, in means and in medians. The days a liquid stock has the most short-marked volume are the days it goes up. The lowest-ratio days are the down days. "Shorting it down" is the inverse of what the file records: on a liquid stock's up day, more of the volume is someone selling shares they do not own to someone who wants them.

The same ordering holds at $10-100M a day (-7.8 bp to +10.6 bp). Below $1M it is louder and noisier - the eighth decile averages +164 bp on the day - and on days a sub-$1M stock rises more than 5% its ratio averages 47.3%, against 40.3% on days it falls more than 5%. The pumps are where the short marks are. That is market makers selling into retail buying, which is what a pump is made of.

Finding 2: in liquid names, the next week goes the other way

Now the registered test. Same sort, same names, return over the next five sessions.

Horizon, >$100M/day	Top minus bottom decile	t	Both halves	Hit rate
Next session	-12.4 bp	-6.4	-10.2 / -14.7	41.8%
Next 5 sessions	-26.7 bp	-4.1	-23.4 / -30.2	39.7%
Skip a day, then 4 sessions	-14.0 bp	-2.5	-12.9 / -15.1	41.5%
Of which, the overnight gap into the next session	-9.9 bp	-9.9	-8.8 / -11.1

By year, the five-session spread is -40 bp in 2022, -9 in 2023, -18 in 2024, -40 in 2025 and -26 in 2026 to date. It is never positive. The deciles are close to monotone the whole way: 34.9, 29.6, 23.8, 19.6, 20.0, 16.2, 10.8, 12.3, 9.7, 7.0 bp from the lowest ratio to the highest. The stock that had the most short-marked volume today gives back, on average, most of what it did today over the following week, and a disproportionate share of that is in the next morning's gap.

This is the academic result, reproduced on 2022-2026 data with a public file. Diether, Lee and Werner showed in 2009 that daily short-sale flow is contrarian - short sellers are active after price rises and ahead of mild reversals. The screeners took the same series and relabelled it a squeeze indicator. The registered bar for a tradable result was a spread with the same sign in both halves and |t| above 2. The >$100M tier clears that bar, on the lore's wrong side, at one day and at five. We claim nothing about a strategy: a 27 bp weekly spread on a ten-decile sort is inside the cost of running it. The claim is the sign.

One tier down, $10-100M a day, the five-session spread is -13.5 bp (t = -2.4) and the next-day spread -6.2 bp (t = -4.0). At $1-10M it is zero in every horizon (+1.1 bp, t = 0.15, over five sessions).

Finding 3: the lore's sign lives below $1M a day, and it is the closing print

The gradient across tiers is the chart we would keep.

In stocks trading under $1M a day - 1,945,795 stock-days - the top-minus-bottom spread over five sessions is +36.2 bp (t = 5.7), positive in both halves (+45.4 and +26.7). On its face, the lore. Three things about it:

• It is the next morning. The next-day spread is +15.2 bp, and the next day's open-to-close spread is +23.7 bp (t = 12.9), while the overnight gap goes the other way (-8.7 bp). Every decile but the top two loses 14-26 bp from the next open to the next close - the standard microcap intraday bleed - and the top decile does not. A microcap that printed heavy short-marked volume today tends to have closed on the bid, and the first hour tomorrow walks it back up. Skip that day and the five-session spread is +20.8 bp (t = 3.3) in the full sample and +13.1 bp (t = 1.2) in the second half.
• It is fading. By year: +49 bp (t = 6.4) in 2022, +41 (5.4) in 2023, +51 (2.8) in 2024, +17 (1.4) in 2025, +19 (0.9) in 2026.
• It is smaller than the spread. We measured the quoted NBBO in this population for the halt ladder: a median half-spread of 0.90% each way. A 36 bp expected move across a week in a name that costs 180 bp to round-trip is not an edge. It is a description of how microcap closing prints behave.

So the one place the screener's arrow points the right way is the one place it cannot be used, and it is the mechanics of the closing print rather than anything about short sellers.

The 50% line, the baseline and the other readings

We also tested the lore as it is actually applied, in the liquid universe (at least $1M a day).

Reading, stocks ≥ $1M/day, next 5 sessions	Spread	t
Ratio above 50% vs at or below	-6.0 bp	-2.2
Ratio above the stock's own 60-session average vs below	-1.5 bp	-0.9
Top decile minus bottom decile	-8.3 bp	-1.6
Top minus bottom, skip a day then 4 sessions	-3.7 bp	-0.8
Top minus bottom, next session only	-4.7 bp	-3.4

Above 50% is worth 4.2 bp over the next week; at or below is worth 4.8. The 52% of stock-days the screener colours red do slightly worse than the 48% it colours green. Using the stock's own history as the threshold - our third registered hypothesis, that deviation from the baseline would carry more than the raw level - does not help: the within-stock reading is indistinguishable from zero. The one version of the lore with any teeth in the liquid universe is the next-day decile spread, and its sign is negative.

What the file can and cannot see

Last, the denominator. The FINRA file records trades reported to a FINRA facility: off-exchange prints from the dark pools and the wholesalers. It does not contain a single exchange print. The "total volume" the screener divides by is not the day's volume. It is the off-exchange share of it.

Measured against the consolidated tape, the file's total covers a median 35% of a stock-day's volume across the panel: 32% in stocks trading at least $1M a day and 47% below $1M, where more of the trading is retail routed through wholesalers. The coverage has risen every year, from 29% in 2022 to 39% in 2025, as off-exchange share has grown - the same trend we measured from the other side in what the TRF tape shows and what it hides. The ratio is a ratio inside the off-exchange tape. A stock whose off-exchange prints are half short-marked and whose exchange prints are a quarter short-marked has a "short volume ratio" of 50% and an actual short share of volume closer to a third. The exchanges publish their own short-sale files, by venue, and nobody adds them up, because the number would stop being alarming.

Two smaller things from the file itself. Short-exempt volume - the sales marked exempt from the uptick rule - is 3.6% of short volume on 14 October 2026 against 1.8% on the same date in 2023. And from 23 February 2026 the file began carrying fractional share quantities: 80% of 2026 rows have a non-integer short or total volume, against none before. The fractions are retail - no institution trades 0.823483 of a share - and they are a small, new, public tracer of where retail orders land.

What we would actually take from this

• The ratio is a liquidity reading, not a positioning reading. It rises when buyers show up and someone sells to them short. Short interest - the bi-monthly count of shares actually borrowed and sold - is a different number from a different source, and the file does not approximate it.
• Above 50% is the median. Any screen keyed on that line is colouring half the market red on half the days.
• In liquid names, read the sign backwards. The highest-ratio decile underperforms the lowest by about 27 bp over the following week, at t = -4, every year. That is contrarian liquidity provision doing what it does. It is not big enough to trade across a spread, but it is big enough to stop treating the number as bullish.
• In microcaps, it is the closing print. The positive spread is the next morning's rebound in names that cost more to cross than the whole effect, and it has halved since 2024.
• The denominator is a third of the tape. Any claim about "x% of today's volume was short" that comes from this file is a claim about the off-exchange third of it.
• If you want the retail footprint, this is not the sharpest tool. We have measured where retail prints actually live in the comb in the fourth decimal - the sub-penny increments that only an internalised order can carry. The short mark and the fraction are blunter versions of the same trace.

Method and caveats

Sources. FINRA Consolidated NMS daily short-sale-volume files (CNMSshvol), every session 3 January 2022 to 14 October 2026, read as published. Daily consolidated bars from the SIP, split-adjusted for returns; dollar volume for the tiers is price times volume, which is invariant to the adjustment; the coverage ratios use unadjusted volume so that a reverse split does not distort them. Security types from the Nasdaq symbol directory, the broker asset master (active and inactive) and the SEC registrant list; 1,499 symbols with no record in any of the three were excluded, almost all of them delisted or sub-$1 names.

Pre-registration. Five hypotheses, the decile sort, the four tiers, the two calendar halves, Newey-West inference with the lag matched to the horizon, and the tradability bar were written down and hashed before any return was joined. H1 (lore: top decile outperforms) is rejected in every tier above $1M a day. H2 (the ratio is a per-stock fingerprint, >50% of variance) is rejected: 15.3%. H3 (deviation from the stock's own baseline carries more than the level) is rejected: it carries less. H4 (high ratio on down days) is rejected: the ratio is highest on up days. H5 (coverage) is reported as measured. Two returns were added after the first tabulation and are labelled as such: the skip-a-day return and the open-to-close return, both to diagnose the microcap result.

What the numbers are not. Decile means are gross of all costs, equal-weight, and include the bid-ask bounce in closing prints; the skip-a-day row exists to show how much of each result is that bounce. A 27 bp weekly spread in liquid names is a sign, not a strategy. The microcap tier includes names that cost multiples of the measured effect to trade.

Independence. Stock-days are not independent within a session; every t-statistic is computed on the time series of daily spreads, so cross-sectional correlation is absorbed, and overlapping five-session returns use a five-lag Newey-West correction.

Survivorship. The panel includes every symbol that appears in the FINRA file and has a bar on the tape, delisted names included, because neither source drops a name after it is gone. The 1,499 unclassifiable symbols are the residual risk, and they are concentrated in the sub-$1M tier, which is not where any conclusion here rests.

The single-stock chart is an illustration. GameStop was chosen because it is the stock the ratio is most often quoted for, not because it is representative of anything.

Nothing here is a recommendation to trade any security. It is a measurement of what one widely sold number has and has not predicted over 1,200 sessions.