Does Daily Short Volume Predict Short Interest? Evidence From 660,000 Settlement Windows

Abstract

Using 660,246 settlement windows across 6,959 stocks from January 2020 through July 2026, this study tests whether FINRA's daily off-exchange short-volume data contains information about concurrent changes in twice-monthly reported short interest. The conventional change in the short-volume ratio produces a pooled Spearman correlation of +0.229.

Among 32 transparent alternatives, an equal-weight composite of within-date ranks for ratio change, market-relative level, and excess short flow raises the pooled correlation to +0.263. In the retrospective 2025–July 2026 test period, raw-feature ridge reaches +0.343 and a 26,213-parameter numeric-transformer ensemble reaches +0.414; the transformer estimates settlement-window endpoints rather than daily positions.

The evidence supports a nowcasting interpretation only. The signals are contemporaneous, have little relationship with the following settlement window, and do not produce a material return forecast. Because the later period was inspected during model development and the FINRA history uses final-file rather than point-in-time vintages, prospective validation remains necessary.

Background and research question

Short interest is a stock—the number of shares held short at a settlement date—whereas short volume is a flow of executions during a trading period. The two measures can overlap, but they are not interchangeable. A market maker, for example, may sell short while filling a customer order and cover later that day; the opening trade appears in the daily short-volume file even though no position remains at the short-interest settlement date.

FINRA accordingly warns that its daily files cover publicly disseminated trades rather than consolidated exchange activity or open positions. The empirical question is therefore narrower than whether short volume “becomes” short interest: does noisy daily flow, aggregated over a complete settlement window, retain information about the contemporaneous change in the reported position?

Data and sample construction

The analysis joins three production datasets. Daily short volume: FINRA's consolidated NMS and OTC daily short-sale files from January 2020 through July 2026, aggregated across the facilities represented in those files. Short interest: twice-monthly observations aligned to FINRA's official settlement and publication calendar. Prices: the exact listed security's split-adjusted daily series for the return tests; securities are never joined through fuzzy ticker matching.

For each stock and settlement date, the trading window begins after the previous settlement date and ends on the current settlement date. Construction follows the official calendar through the May 2024 transition from T+2 to T+1 settlement.

Sample filters

The final eligible sample contains 660,246 windows across 6,959 stocks and 156 settlement dates.

The comparable formula panel contains 653,459 observations for which every component of the baseline and composite scores is available. The difference is primarily the initial history required to calculate changes and trailing expectations.

Short interest increased in 49.7% of eligible windows. The cross-sectional average short-volume ratio was 46.4% of the off-exchange volume in the files.

Empirical design

For stock i in settlement window t, define the window's short-volume ratio as:

R_{i,t} = \frac{\text{total short volume in window }t}{\text{total reported volume in window }t}

The ratio is therefore volume-weighted across the days of the window.

The outcome is:

\Delta SI_{i,t} = \frac{SI_{i,t} - SI_{i,t-1}}{SI_{i,t-1}}

Association is measured with Spearman rank correlation because short-interest percentage changes have extreme tails and the question is ordinal: whether higher signal values correspond to larger position changes. We report both a pooled statistic across all stock-windows and the mean of correlations calculated separately for each settlement date, which prevents dates with larger cross-sections from dominating. Newey-West standard errors with six lags are used for time-series inference.

All cross-sectional ranks and quintiles are formed within each settlement date; observations from different dates are never ranked against one another.

Baseline results

The conventional predictor is the change in the window's short-volume ratio:

\Delta R_{i,t} = R_{i,t} - R_{i,t-1}

Its pooled Spearman correlation with the concurrent short-interest change is +0.229.

A second predictor compares the current ratio with the market's equal-weight average on the same settlement date:

R^{\mathrm{relative}}_{i,t} = R_{i,t} - \overline{R}_{\mathrm{market},t}

Its pooled correlation is +0.190.

Averaged across settlement dates, the two correlations are +0.223 and +0.185. Their Newey-West t-statistics are approximately 32 and 46.

The market-relative result is positive on all 156 settlement dates.

Among the 5,012 stocks with enough observations for a stock-specific estimate, the ratio-change relationship is positive for 97.6%. The median within-stock correlation is approximately +0.23.

This is not merely a comparison between permanently high-ratio and permanently low-ratio stocks. Demeaning each stock's history leaves correlations of +0.219 for the ratio change and +0.193 for the market-relative ratio.

Probability that short interest increased

The crossover occurs near the market's average ratio rather than at the psychologically intuitive 50% threshold.

That is why a market-relative measure is more meaningful than declaring every ratio above 50% "high."

Lead-lag analysis

To distinguish a contemporaneous signal from persistent differences across stocks, each measure is aligned with the preceding, concurrent, and following short-interest changes.

Both measures peak in their own settlement window. Ratio change is negatively associated with the prior window, consistent with some mean reversion, while both next-window correlations are approximately zero. A separate placebo for the three-factor composite below also produces a small negative next-window correlation of roughly -0.01.

The timing supports a nowcast: daily flow helps estimate the short-interest change forming inside the current settlement window before that report becomes public. Once the next window begins, the old signal supplies almost no directional information about the new position change.

Alternative specifications

The baseline ratio treats every short-volume share alike.

That is unlikely to be optimal. A 60% ratio on unusually heavy volume may contain more information than the same ratio on a quiet tape. A stock already accustomed to a 60% ratio may also be less unusual than one whose ratio has abruptly moved from 40% to 60%.

We tested 32 fixed formula variants. The search covered ratio levels and changes, short volume excluding short-exempt volume, recent-day concentration, and rolling means, medians, and exponentially weighted expectations. It also tested excess short shares relative to an expected ratio, scaling by prior short interest, abnormal-volume interactions, within-date rank composites, and chronological ridge models using only information available by the window close.

This was an exploratory formula search. The historical periods help test stability, but they were not an untouched, preregistered experiment.

Selected formula results

Several intuitive modifications fail to improve the result. Short-exempt volume represents only about 1.33% of recorded short volume, and subtracting it slightly reduces correlation. Restricting the calculation to the final three trading days performs substantially worse, indicating that the information is distributed across the settlement window. Additional short-horizon features sometimes help in the latest period but weaken the 2024 validation result.

Transparent three-factor specification

The strongest formula that remains simple, auditable, and free of fitted coefficients has three components.

1. Change in the ratio

\operatorname{Change}_{i,t} = R_{i,t} - R_{i,t-1}

This measures whether short-sale flow accelerated or decelerated relative to the previous settlement window.

2. Market-relative level

\operatorname{Relative}_{i,t} = R_{i,t} - \overline{R}_{\mathrm{market},t}

This measures whether the stock's current ratio is high or low compared with other securities facing the same market regime.

3. Excess short flow relative to prior short interest

First calculate the stock's expected ratio from as many as four previous windows:

\widehat{R}_{i,t} = \operatorname{mean}\!\left(R_{i,t-1}, \ldots, R_{i,t-K}\right),\quad K \le 4

Then calculate:

\operatorname{Excess}_{i,t} = \frac{SV_{i,t} - \widehat{R}_{i,t} V_{i,t}}{SI_{i,t-1}}

The numerator measures short-volume shares above or below the amount implied by the stock's recent ratio at current volume. Dividing by prior short interest scales that excess to the previously reported position. The construction does not assume those shares remain open; it gives unusually large flow an economically comparable scale.

Combining the components

Within each settlement date, convert each component to a percentile rank.

The final score is:

\operatorname{Score}_{i,t} = \frac{\operatorname{Rank}(\operatorname{Change}_{i,t}) + \operatorname{Rank}(\operatorname{Relative}_{i,t}) + \operatorname{Rank}(\operatorname{Excess}_{i,t})}{3}

The score has no trained coefficients. Each component receives equal weight, and the result remains between zero and one; a value near one indicates high ranks on all three dimensions relative to other stocks in the same settlement window.

Worked example

Suppose a stock has a current short-volume ratio of 58%, a prior-window ratio of 47%, a current market average of 46%, and a trailing expected ratio of 45%. It traded two million shares of total off-exchange volume, while previously reported short interest was four million shares.

The ratio change is 58% − 47%, or +11 percentage points. The market-relative ratio is 58% − 46%, or +12 percentage points. Excess short flow is (1.16 million − 0.90 million) / 4 million, or 6.5%.

If those values rank at the 88th, 91st, and 84th percentiles on that settlement date, the final score is:

\frac{0.88 + 0.91 + 0.84}{3} = 0.877

That does not mean short interest increased by 87.7%. It means the stock has a relatively strong nowcast compared with other stocks in the same window.

Chronological stability

The full-history pooled correlation rises from +0.2289 to +0.2630, an absolute improvement of 0.0341 and a relative improvement of 14.9%.

Pooled correlations by chronological period

The improvement is present in every chronological period. The baseline itself strengthens from +0.202 in development to +0.275 in the retrospective test, so part of the composite's higher recent correlation reflects a broader increase in signal strength.

Mean per-settlement-date correlations

The date-by-date estimates agree with the pooled results: the three-factor formula improves the baseline without fitted coefficients, while the expanded ridge extracts more historical correlation at the cost of greater model-selection risk.

Component interpretation

The three inputs capture different aspects of the same settlement window. Ratio change measures acceleration relative to the stock's immediately preceding window; the market-relative level separates a stock-specific elevation from a broad market shift; and excess flow asks whether the deviation occurred on enough volume to be large relative to the previously reported short position.

The measures overlap without being redundant. A stock may have a high ratio but little change from its own history, a large increase while remaining below the market, or a high ratio on too little volume to matter relative to its existing short base. Averaging their within-date ranks rewards observations that are strong across these complementary dimensions.

Abnormal-volume interaction

The excess-flow component is consistent with a separate result from the original analysis: short-volume ratios are more informative when the stock trades heavily relative to its own recent history.

We define abnormal volume as:

\operatorname{AbnormalVolume}_{i,t} = \frac{\operatorname{AvgDailyVolume}_{i,t}}{\operatorname{median}\!\left(\operatorname{AvgDailyVolume}_{i,t-1}, \ldots, \operatorname{AvgDailyVolume}_{i,t-K}\right)},\quad K \le 4

A value of 2.0 means the stock traded at twice its own recent norm. In a per-date regression of short-interest change on the short-volume ratio, abnormal volume, and their interaction, the standardized interaction coefficient is +0.037 with a Newey-West t-statistic of 13.0. Using raw volume instead produces an interaction of only +0.008, making company size a poor substitute for whether the stock's own activity is unusual.

Correlation by abnormal-volume quintile

The all-stock sample contains one dip in the second quintile, while the liquid-stock sequence is strictly monotonic. The interaction approximately doubles from +0.026 in 2020 to +0.059 in 2026, with a temporary dip in 2022, and falls to approximately zero against the following window's short-interest change.

Probability that short interest rose

The probability spread between the lowest and highest ratio quintiles increases from 21 percentage points on the quietest tape to almost 28 points on the hottest. Some amplification may be mechanical because ratios estimated from thin volume contain more measurement error; the remainder may reflect genuine information in unusually heavy short-sale flow. These data cannot cleanly separate the two explanations.

Earnings-window robustness

Because unusually high volume clusters around earnings announcements, earnings windows are a potential confound. We identify such windows among the 3,654 stocks with sufficiently complete call calendars, covering 59.7% of the full panel. Median abnormal volume is 1.28 during earnings windows and 0.95 outside them.

Despite that difference in activity, the short-volume correlation is similar in the two groups: +0.182 during earnings windows and +0.177 otherwise. The abnormal-volume result is therefore not explained solely by the earnings calendar.

Chronological ridge benchmark

A chronological ridge model provides a fitted benchmark for the transparent formulas. Every input is restricted to information available by the settlement-window close. Direct flow features include ratio change, market-relative ratio, four-window excess short flow, unweighted daily ratio change, final-three-day ratio change, the first-half versus second-half path, and volume-weighted ratio change. Context features include abnormal volume, turnover relative to prior short interest and prior FINRA volume, previously reported days to cover, daily ratio dispersion and slope, and selected interactions.

Present-day market capitalization and shares outstanding are excluded because current database values would introduce historical look-ahead. The model is trained through 2023, its regularization strength is selected on 2024, and the resulting specification is evaluated retrospectively from 2025 through July 2026. The largest standardized contributions come from excess short flow, abnormal volume, unweighted daily ratio change, turnover relative to prior short interest, baseline ratio change, and the market-relative ratio.

The expanded model reaches +0.3107 pooled Spearman over the complete history, +0.3072 mean per-date, and +0.3539 mean per-date in the retrospective 2025–July 2026 period. Two independent data extractions reproduce the broad result.

This fitted estimate is not the headline result. Although the inputs are point-in-time, feature families were explored after inspecting the historical data, so the research process is not a pristine out-of-sample experiment. The ridge result measures how much structure may remain in the daily data; it is not yet a confirmed live correlation of +0.31.

Return tests

The concurrent short-volume signal has a small positive relationship with the current window's return, approximately +0.06. This sign is consistent with several mechanisms that the aggregate files cannot separate: market makers may sell short while filling customer buying demand, and informed short sellers may choose to sell into strength rather than chase falling prices.

The original pre-specified tests show essentially no next-window return predictiveness.

One narrower result is statistically detectable. Within the highest abnormal-volume quintile, a high market-relative ratio correlates at -0.036 with the following window's return, with a Newey-West t-statistic of -7.6. The comparable short-interest correlation is -0.048, and the relationship appears both inside and outside earnings windows.

A two-week rank correlation near 0.04 is economically small and occurs where trading costs matter most, so it is better interpreted as a microstructure result than as a strategy. The three-factor formula was selected to nowcast short interest, not returns; any post-selection return test requires independent validation.

Robustness checks

The baseline and composite results survive several changes to the construction.

Calendar alignment. Shifting the trading window for the settlement convention raises the baseline ratio-change correlation to approximately +0.249 rather than eliminating it.

Calendar years. The baseline is positive in every year, ranging from approximately +0.206 in 2020 to +0.278 in 2026.

Within-stock variation. Demeaning each stock's series leaves the baseline relationships close to their pooled values.

Outcome tails. Winsorizing extreme short-interest changes does not remove the result.

Liquidity. Restricting the panel to the 1,500 most active stocks preserves the relationship.

Short-exempt volume. Subtracting short-exempt shares slightly weakens rather than improves the baseline.

Splits. Every window containing a recorded split is excluded.

Identity matching. Price series are joined by the exact listed security rather than ticker similarity.

Endpoint reconciliation. The previous-position value carried in each short-interest release agrees with the preceding stored observation for 98.85% of 855,221 comparable pairs. Most remaining differences are within 2%, consistent with revisions.

Return coverage. Approximately 2.9% of eligible windows lack a same-window return and 3.9% lack a following-window return.

The missing forward returns include delistings, so return conclusions retain mild survivorship bias. This limitation matters more for return analysis than for the concurrent short-interest result.

Numeric-transformer specification

The transparent formula was designed to remain simple and auditable. We then tested whether a learned model could use the shape of the daily path without treating market values as words or inventing an unobservable daily short-interest label.

The result is a 26,213-parameter causal numeric transformer. Each sequence position represents one FINRA trading day and carries 21 continuous or availability features. They describe the daily short ratio, non-exempt and exempt flow, total and short volume, volume relative to 21-day and 63-day history, ratio changes, market-relative activity, flow relative to previously reported short interest, turnover relative to filed shares outstanding, prior short-interest behavior, days to cover, timing, and missingness.

A second input contains 14 raw settlement-window aggregates and 14 matching availability masks. These cover ratio change and level, excess short flow, recent-day concentration, slope and dispersion, abnormal volume, reported-volume coverage, prior short-interest change, days to cover, and window length. The learned inputs contain no rank calculated over the stocks that eventually reported the target.

The network projects each day to 32 dimensions, applies two causal self-attention layers with four heads and a 96-unit feed-forward block, and combines the daily representation with the aggregate branch. Separate heads produce the median log change, P10 and P90 bounds, and an auxiliary uncalibrated rank score.

Chronological training protocol

We trained three deterministic runs with seeds 17, 29, and 43. Each used AdamW, batches of 512, a learning rate of 0.0005, weight decay of 0.0001, gradient clipping at 1.0, and at most eight epochs. Early stopping selected the epoch with the strongest mean per-settlement-date Spearman correlation in 2024.

The primary loss was Huber loss on the endpoint log change, with delta 0.15. Two auxiliary terms each received a weight of 0.2: P10/P90 pinball loss for the interval and mean-squared error on the within-date target rank. Dropout was 0.1.

Only after every validation gate passed did we train the released checkpoint from scratch on all 660,246 labeled windows. That full-history fit used seed 17 and seven epochs, the median selected epoch across the three validation runs. The released checkpoint itself has no honest out-of-sample score because every labeled row was used to fit it; the table below belongs to the frozen retrospective evaluation ensemble.

Retrospective model comparison

The transformer’s median absolute log error is 0.0928, direction accuracy is 63.4%, and its P10–P90 interval covers 82.4% of outcomes. Individual test correlations are +0.4079, +0.4048, and +0.4040, with a standard deviation of 0.00171. The original +0.229 result refers to the full-history panel; the appropriate baseline for the later retrospective test is +0.2751 versus +0.4143 for the ensemble.

Released model artifacts

The trained weights are available at daniel3303/equibles-short-interest-nowcast on Hugging Face under Apache-2.0.

The repository contains an executable ONNX graph, safetensors weights, normalization metadata, the exact input and output contract, checksums, hyperparameters, and the full model card. It does not contain the production snapshot, processed training examples, or private training source.

The ONNX model accepts a [batch, 32, 21] daily tensor, a [batch, 28] aggregate tensor, and a [batch, 32] valid-day mask. Its four outputs are P50, P10, and P90 endpoint log changes plus the auxiliary rank score. To reconstruct an endpoint estimate from any log-change output, add it to log1p of the previously reported short interest and apply the inverse transform.

Retrospective-evaluation limitations

The historical FINRA store keeps the latest imported value for each partition rather than every publication vintage. All 1,641 partitions used through the test cutoff were imported on 6 August 2026. This is therefore chronological retrospective final-file evaluation, not strict point-in-time or zero-look-ahead evidence.

The 2025–July 2026 period had also been inspected during earlier formula research. A genuine forward test must archive each prediction before the corresponding official short-interest value becomes public. Until that happens, +0.414 should be read as a strong retrospective result rather than a promised live correlation.

Interpretation and scope

The composite can rank stocks by the likelihood and relative strength of the short-interest change currently being formed. It may help monitor whether reported short interest is likely rising or falling before publication, prioritize securities for deeper borrow, options, or positioning research, identify high-ratio activity that is also unusual for the stock and large relative to its prior short base, and compare stocks on the same settlement date without using a universal 50% threshold.

It cannot reveal how many short-volume shares remain open or who initiated the trades. It cannot distinguish market making, hedging, arbitrage, and directional shorting; supply the exact short-interest value that will be published; predict whether short interest will continue moving in the next settlement window; or establish that the stock's price will subsequently fall.

A score near one means "strong relative evidence of a concurrent increase," not "nearly all trades created new short positions."

Reproducible specification

The calculation is mechanical.

1. Build consecutive short-interest settlement windows for each exact security.
2. Aggregate FINRA short and total volume between the two settlement dates.
3. Calculate the volume-weighted short-volume ratio.
4. Calculate the change from the previous window.
5. Subtract the current date's equal-weight market average.
6. Estimate the stock's expected ratio from up to four prior windows.
7. Convert the difference between actual and expected short volume into shares.
8. Divide those excess shares by the previously reported short interest.
9. Percentile-rank the three components within the settlement date.
10. Average the three ranks.
11. Compare the score with the concurrent percentage change in reported short interest using Spearman correlation.

The primary reproducible specification should remain frozen as:

\operatorname{Score}_{i,t} = \operatorname{mean}\!\left[\operatorname{Rank}(\Delta R_{i,t}),\operatorname{Rank}(R^{\mathrm{relative}}_{i,t}),\operatorname{Rank}(\operatorname{Excess}_{i,t})\right]

No coefficients, market-cap data, current shares outstanding, or future observations enter the formula.

Prospective validation protocol

Historical date splits cannot create a genuinely untouched test after the formulas and feature families have already been examined. The prospective test therefore begins after 15 July 2026, with the specification frozen before additional short-interest releases are observed.

The benchmark is the change in the window short-volume ratio; the primary transparent model is the equal-weight three-factor rank composite; and the learned challenger is the published numeric-transformer checkpoint with its documented feature contract. Pooled Spearman correlation with the concurrent short-interest change is the primary metric, mean per-settlement-date Spearman is secondary, and correlation with the following window is the placebo. Results should be reported for both the full sample and the 1,500 most liquid securities.

No formula or filter changes should be made during the test. The first formal evaluation is scheduled after approximately 24 new settlement dates. The expanded ridge may be frozen as a separate challenger but should not replace the transparent specification unless it succeeds on observations that played no role in model selection.

Conclusion

Daily short volume and short interest measure different objects: executed short-sale flow versus positions remaining open at a settlement date. The distinction emphasized by FINRA is essential, but the two series are not empirically unrelated.

Across more than 660,000 settlement windows, the conventional change in short-volume ratio has a +0.229 pooled Spearman correlation with the concurrent short-interest change. A transparent composite of ratio change, market-relative level, and excess short flow raises the correlation to +0.263 without fitted coefficients. In the retrospective 2025–July 2026 test, the numeric-transformer ensemble reaches +0.414, compared with +0.343 for ridge and +0.305 for the transparent composite.

The evidence supports using daily short-volume data to nowcast the direction of an unpublished short-interest report. It does not identify the exact open position, predict continued accumulation, or establish a standalone return signal. The published weights make the learned specification reproducible and freezeable; confidence comparable to a preregistered test will require the prospective evaluation described above.

References

• FINRA, Understanding Short Sale Volume Data on FINRA's Website.
• FINRA, Daily Short Sale Volume Files.
• FINRA, Short Interest Reporting.
• Boehmer, E., Jones, C. M., and Zhang, X. (2008), Which Shorts Are Informed?, Journal of Finance, 63(2), 491–527.
• Diether, K. B., Lee, K.-H., and Werner, I. M. (2009), Short-Sale Strategies and Return Predictability, Review of Financial Studies, 22(2), 575–607.