How accurate is the short interest forecast?
We tested the previous and current models on January 2025 through July 2026. Neither model trained on any row from that period.
The current forecast was more accurate on every main measure:
| Measure | Previous model | Current model | Change |
|---|---|---|---|
| Directional accuracy | 58.3% | 61.6% | +3.3 points |
| Rank correlation | +0.3235 | +0.3498 | +0.0263 |
| Rank correlation by settlement date | +0.3208 | +0.3435 | +0.0227 |
| Typical sizing error | 0.1014 | 0.0967 | smaller |
| Forecast range coverage | 80.5% | 81.2% | +0.7 points |
| Rows tested | 193,257 | 196,195 | newer snapshot |
Directional accuracy is the easiest number to read: the model correctly forecast whether reported short interest would rise or fall 61.6% of the time.
Rank correlation asks whether the model put stocks with the largest increases above stocks with the largest decreases. Higher is better. The current model scored +0.3498.
The range around each forecast contained the final FINRA number 81.2% of the time. That range is intentionally wide because the exact position is still unknown.
What the forecast predicts
FINRA publishes short interest twice a month, usually about two weeks after the settlement date. Short interest is the number of shares investors held short on that date.
FINRA also publishes short-sale volume every trading day. Short-sale volume counts trades, not open positions, so it is not the same as short interest.
Our earlier study, Does Daily Short Volume Predict Short Interest?, found that daily short volume contains useful information about the position building inside a settlement window.
The Equibles forecast turns those daily files into four reader-facing answers:
- The estimated short position at the next settlement date.
- Whether short interest is expected to rise or fall.
- A likely range around the estimate.
- A confidence level for the direction.
The challenge is timing. A forecast made on day two of a settlement window has much less evidence than one made on day ten. The results in this article average across that full range because that is how readers use the product.
Why confidence matters
The model estimates the probability that short interest will rise. That probability separates weak calls from stronger ones.
We chose the confidence thresholds using 2024 data and then applied them unchanged to the later test period:
| Forecast group | 2024 accuracy | 2025 to July 2026 accuracy |
|---|---|---|
| All forecasts | 61.7% | 61.6% |
| Most confident half | 68.6% | 68.8% |
| Most confident third | 72.0% | 72.3% |
The close match between the two periods matters. It shows that the confidence thresholds continued to work after they were set.
For the most confident third of forecasts, the direction was correct about 72% of the time. A random guess would be correct 50% of the time.
What changed from the previous model
The current short interest forecast adds two things:
- It directly estimates the probability that the reported position will rise.
- It uses the stock's price behavior during the same settlement window alongside daily short-sale data.
Price context matters because unusual short selling can mean something different when a stock is falling sharply than when its price is flat.
Why we corrected the first release
The version released on 1 September 2026 ranked stocks well and called direction well, but its central estimates were too low.
The day after launch, it showed “falling” for 79% of a 190-stock sample. Once the settlement windows were complete, that rose to 93%. No settlement during the previous six months had seen more than 54% of stocks fall.
The training process had not forced the central estimate to stay correctly centered. As more days entered a window, the bias grew from about one percentage point to about five. Three training runs that differed only by random seed forecast declines for 40%, 64% and 65% of rows.
Our original acceptance measures did not expose the problem. Rank correlation does not change when every estimate shifts by a similar amount, and the typical sizing error changes very little.
The corrected model now measures and removes that bias separately for each number of observed days. We also reduced the weight placed on the direction task during training.
Three new checks must now pass on data the model did not train on:
- The typical signed error must stay within 1.5 percentage points of zero.
- The share of forecasts showing “rising” must stay within ten points of the actual share.
- Different random seeds must not produce materially different forecast levels.
The corrected version improved ranking and sizing while keeping directional accuracy and range coverage nearly unchanged:
| Measure | 1 September version | Corrected version |
|---|---|---|
| Rank correlation | +0.3395 | +0.3498 |
| Typical sizing error | 0.0984 | 0.0967 |
| Directional accuracy | 61.8% | 61.6% |
| Forecast range coverage | 81.6% | 81.2% |
| Estimate and direction agree | 75% | 89% |
Everything else in this article refers to the corrected forecast.
Why this test is harder than the earlier study
Our earlier article reported a model score of +0.4143. That test used completed settlement windows, after every daily file in the window was available.
The live product has to forecast short interest before a window is complete. On this harder test, the previous model scored +0.3235 with 58.3% directional accuracy. Those are the relevant numbers for comparison.
When the current model is tested on completed windows like the earlier study, it reaches +0.4274 rank correlation and 64.1% directional accuracy.
Data and testing
| Data item | Coverage |
|---|---|
| Full data period | January 2020 to August 2026 |
| Settlement-window observations | 682,982 |
| Stocks | 7,353 |
| Settlement dates | 158 |
| Daily short-sale files | 2 January 2020 to 1 September 2026 |
| Snapshot date | 2 September 2026 |
Each observation represents one stock and one settlement window as seen on a particular day inside that window. It uses only the daily files published by that point.
A row qualifies when it has at least one day of short-sale data, a previous short position of at least 10,000 shares, average daily off-exchange volume of at least 10,000 shares and no stock split during the window. Stocks are matched by exact listed identity, never by ticker similarity.
We trained on data through 2023, selected settings and stopping points with 2024 data, and tested on January 2025 through July 2026. Settlement dates never cross from one split into another.
Comparison with simple benchmarks
The same held-out period and eligibility rules were used for every method:
| Method | Rank correlation |
|---|---|
| Change in short-volume ratio | +0.2053 |
| Three-factor rank composite | +0.2154 |
| Ridge regression on raw features | +0.2508 |
| Previous model | +0.3235 |
| Current model | +0.3498 |
What the forecast cannot tell you
- It cannot give the exact next position. It provides an estimate and a range. The range contained the reported value about 81% of the time.
- It does not predict stock prices. Our earlier study found almost no ability to predict the next settlement window's price return.
- It is not a live trading record. Historical FINRA daily files are stored in their latest available form, so later file revisions can appear in this study.
- The test period was not new to us. We examined 2025 and 2026 while researching the earlier formulas. The data was held out from model training, but not from the researchers.
Why live-model scores look different
The results above come from models tested on data they did not train on. Those are the figures readers should use to judge forecast accuracy.
The production model is refitted on every available row. Scores measured on those training rows are optimistic and do not estimate future performance. We include them only to show how the low-bias problem changed across releases:
| Measure | Previous release | 1 September release | Current release |
|---|---|---|---|
| Rank correlation | +0.3239 | +0.3475 | +0.3649 |
| Rank correlation by date | +0.3195 | +0.3530 | +0.3586 |
| Directional accuracy | 60.4% | 63.1% | 62.8% |
| Typical sizing error | 0.0972 | 0.1038 | 0.0956 |
| Typical signed error | +0.001 | −0.035 | −0.005 |
| Forecasts showing “rising” | 57.6% | 31.7% | 51.5% |
| Positions that actually rose | 51.7% | 51.7% | 51.7% |
The current release is centered close to the market. The 1 September release was not.
Why the current model is closed
We published the previous model's weights, runnable graph and complete input contract on Hugging Face under Apache-2.0. They remain available for anyone who wants to reproduce or improve on the +0.3235 result.
We are not releasing the current model's weights or design.
Instead, we are publishing the evidence that can be checked: the test method, the held-out results and the ongoing forecasts. Twice a month, FINRA's official figure provides a new result against which the model can be judged.
The transparent three-factor formula from the earlier study also remains public and unchanged.
Where to find the short interest forecast
Every covered stock with enough short-sale history has a forecast on its Equibles page. It shows the expected position, change from the last official figure, range, direction and confidence.
The REST API exposes the same fields at /v1/stocks/{ticker}/short-interest/forecast: riseProbability, expectsIncrease and confidence. Use expectsIncrease for direction because the model's decision threshold is not 0.5.
Connected assistants receive the same forecast, confidence band and historical range accuracy.
The forward test
Our earlier study set a forward-testing plan: keep the rules unchanged, compare against the short-volume ratio and the transparent three-factor formula, and review the result after about 24 new settlement dates.
That plan remains in place. The corrected model entered the test on 2 September 2026. The version released on 1 September was withdrawn because a one-day model does not provide a meaningful forward record.
Method details
The forecast target is the change in the natural logarithm of reported short interest:
y_{i,t} = \log(1 + SI_{i,t}) - \log(1 + SI_{i,t-1})
Rank correlation is Spearman correlation. We calculate it across all stock-window observations and separately by settlement date so that dates with more stocks cannot dominate the result.
Directional accuracy compares the model's probability, using its fixed threshold, with whether the reported position actually rose. Typical signed error is the median forecast minus outcome in log terms; a negative value means the typical estimate was too low.
Scaling uses training data only. All thresholds, confidence cutoffs and stopping points were selected with 2024 data and applied unchanged to the later test period.