You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This dataset is free to browse and gated for download. Approval is tied to a Papers With Backtest subscription, which also covers the other datasets in this organisation and the strategy catalogue at https://paperswithbacktest.com. Plans and what each one includes: https://paperswithbacktest.com/pricing

Log in or Sign Up to review the conditions and access this dataset content.

Stocks Weekly ShortOverShorted

Weekly short interest of US stocks, with the share count, market capitalisation and volume beside it, on one grid of Fridays since 1998.

3,789,249 rows over 6,745 symbols, 12 columns, covering 1998-01-02 to 2026-05-29. The most recent observation is 2026-05-29.

Why It Matters

Short interest on its own says little without the size of the company and the liquidity behind it. This dataset carries both alongside it, on one weekly grid:

  • Short interest as of each Friday: short_interest is the latest exchange-reported figure at least 15 days old on the Friday, so it was public by then (FINRA disseminates each settlement-date report about eight business days later).
  • Days to cover, as FINRA defines it: days_to_cover divides that short interest by the average daily volume of its own reporting window, on every row since 1998.
  • Size in the same row: market_cap and number_of_shares give the capitalisation and the share count, so a short-interest ratio or a size control needs no second join; short_percentage ranks that ratio across the week.
  • Liquidity: volume is the week's mean daily volume.

One definition per column. The original feed stopped in June 2026. From 2026-06-26 the series is rebuilt from public sources, checked against the original on the year before: short_interest from FINRA's consolidated short interest (identical to the original on all of the 278,294 rows compared over the year before the freeze, and on 99.5 % of the 1.83 million rows FINRA's public history covers since 2018), number_of_shares from the share count on the cover page of the latest 10-Q or 10-K filed with the SEC, market_cap from the Friday's close, volume from the daily volumes. days_to_cover and short_percentage are computed again on the whole history, with the same rule before and after the freeze, and number_of_shares is on the Friday's own share basis on every row; the original provider's model columns are not published. On the rows from 2026-06-26 a missing value is empty, where the earlier rows wrote a zero (a zero share count or capitalisation is now empty everywhere).

Changes in November 2026

Compared with the previous version of this dataset:

  • Removed: over_shorted, over_shorted_chg and short_prediction, the original provider's model. No public source reproduces it, and it was empty after June 2026.
  • days_to_cover redefined on every row. It is now FINRA's ratio: short interest over the average daily share volume of the reporting window, the trading sessions from the day after the previous settlement date through the settlement date, with the volume on the settlement date's share basis, rounded to hundredths. The previous figure was the provider's own and moved against FINRA's (rank correlation -0.59 over 2018-2026). FINRA's published figure shows 1.00 for any ratio under one; this column keeps the ratio.
  • number_of_shares on one basis. The rows before 2026-06-26 carried the count restated for every later split (NVDA's May 2024 rows showed 24.6 billion shares, the count after the 10-for-1 of June 2024). They are now on the Friday's own basis, like the later rows and like the short interest: about 782,000 rows of some 1,800 stocks change. On that basis market_cap, unchanged, equals the Friday's close as printed times the count on 92.8 % of those earlier rows (75.1 % with the restated count; 83.7 % against 2.8 % on the rows a later split touches), and the count agrees within 5 % with the SEC's cover page on 93.6 % of the rows compared (84.2 % before). A spin-off's price adjustment, which the splits table records like a split, changes no share count and is not applied to it. A zero count or capitalisation, the provider's missing value, is now empty.
  • short_percentage recomputed on every row, with the rule the extension uses: the week's percentile rank of short interest over shares outstanding, among the rows that have both, both on the Friday's basis.
  • Listings since the freeze added: stocks FINRA reports after June 2026 that the catalogue routes to Stocks-Daily-Price, that had no price before the freeze and whose ticker the dataset never held.

How the two derived columns are built

  • Settlement dates: the 15th of each month, or the last session before it, and from September 2007 also the month's last session (monthly only before). This calendar reproduces every one of the 212 settlement dates FINRA publishes since 2017-12-29, and every week the short interest changes in this dataset since 1998.
  • Volume: consolidated daily share volume from Stocks-Daily-Price, put back on the settlement date's share basis with Stocks-Daily-Splits. A session with no print counts as zero, as in FINRA's average. On 2018-2026 this average is within 5 % of FINRA's own on 79 % of stock rows (median ratio 1.002), and the ratio within 5 % of FINRA's on 80 % of the rows FINRA does not floor at one, 91-97 % since 2022. The gap is FINRA's: from December 2019 to May 2021 its average divides by about one session more than the window holds.
  • Depth: both columns start with the dataset, in 1998, wherever the row has a positive short interest and, for days_to_cover, the symbol has volume in the window (95 % of rows), or for short_percentage, a share count (94.5 %; from 1998-12-04, the 1998 rows having almost no share count). FINRA's own public history of short interest begins on 2017-12-29 and the SEC's cover-page share counts in 2009; the earlier rows use the short interest and share counts this dataset already carried, the same quantities (see above).

Load It

Installation/Upgrade:

pip install --upgrade pwb-toolbox

Load the Dataset:

from pwb_toolbox import datasets as pwb_ds

df = pwb_ds.load_dataset("Stocks-Weekly-ShortOverShorted", symbols=["GME"])
print(df.iloc[-1, :])

Columns

Column Name Description
symbol Stock ticker.
datetime End of the reporting week.
short_interest Shares sold short, from the latest settlement date at least 15 days before the Friday.
number_of_shares Shares outstanding, from the latest filing public on the Friday, on the Friday's own share basis (not restated for later splits).
short_percentage The week's percentile rank, between 0 and 1, of short interest over shares outstanding (on the Friday's share basis), among the rows that have both. Not a percentage of the float.
days_to_cover Short interest over the average daily share volume of its reporting window (FINRA's definition, not floored at one), to hundredths. Empty where the short interest is not positive or the window has no volume.
market_cap Market capitalisation, in millions of USD: the Friday's close times number_of_shares.
total_revenue Quarterly revenue, in USD. Until June 2026 interpolated between two reports (so a week already leans toward a figure not yet published); from 2026-06-26, the latest quarter filed by the Friday.
volume Mean daily volume of the week, in shares.

Short interest, share counts, capitalisation, revenue and volume provided by SOV.AI until 2026-06-19; from 2026-06-26, rebuilt from FINRA's consolidated short interest, the SEC's EDGAR filings and the catalogue's daily prices. days_to_cover and short_percentage are computed by Papers With Backtest on every row (FINRA's definition of days to cover).

Access

Browsing the card and the schema is open to anyone. Downloading the files needs an approved request, tied to a subscription: what each plan includes. The same subscription covers the other datasets in this organisation.

Elsewhere

Papers With Backtest publishes 32 datasets on the Hub and codes the papers that use them. Every strategy in the catalogue is run over its own full history before it is published, which is where the numbers above come from.

Downloads last month
296