ZipLime/us-short-interest

US Short Interest

This dataset publishes no short-interest data, and that is the finding, not a bug.

Every US source of equity short positioning was audited on 13 September 2026 before any of it was collected. Two forbid the publication pattern and the third does not exist yet. What is published here is the audit itself — which clause, on which page, read on which day — plus the adapters that will produce data the moment either fact changes.

If you came here for short interest numbers, the last section says where they actually are and what you have to agree to in order to use them.

The finding

Source	Status	Why
FINRA Equity Short Interest	restricted	terms forbid this publication pattern
Exchange feeds (Nasdaq, NYSE)	restricted	subscription products, redistribution rights not obtained
SEC aggregated Form SHO	does not exist yet	compliance exempted until 2 January 2028

FINRA

FINRA's Specific Terms for Equity Data permit redistribution, but only on conditions an Apache-2.0 dataset cannot meet. Three of them, quoted as read:

2.3 Developer may redistribute the Equity Data … to third party end users … for such End Users' non-commercial personal or professional use only

Apache-2.0 grants every recipient the right to use the work commercially.

2.3(c) No Further Redistribution. End Users shall not be permitted to further redistribute the Equity Data received from Developer.

Apache-2.0 grants exactly that right to every recipient. Publishing FINRA data under it would be granting what the licence forbids.

2.3(d) Compliance Efforts. Developer shall take commercially reasonable efforts to ensure that End Users … comply … (for example, by requiring End Users to enter into agreements with Developer that require such compliance).

A public dataset has no such mechanism.

FINRA's website terms of use are narrower still, and they cover the publication schedule as well as the data — content

may not be copied, reproduced, transmitted, displayed, performed, distributed, rented, sublicensed, uploaded, posted, framed, altered, stored for subsequent use, or otherwise used in whole or in part in any manner without FINRA's prior written consent

and shall be used "ONLY for your own non-commercial personal or professional use". Those terms also prohibit using any portion of the site "in conjunction with any machine learning, neural network, deep learning, predictive analytics or other artificial intelligence computer or software program".

So neither the positions nor the calendar of when they are released can be republished here.

Exchanges

Nasdaq publishes short interest by issue to subscribers: "Subscribe to access a comma-delimited text file (.txt) via Secure File Transfer Protocol." A subscription feed carries its own redistribution terms. No subscription was taken out, so no terms were obtained, so nothing was collected.

SEC Form SHO — the open alternative, in 2028

Rule 13f-2 requires managers above certain thresholds to file Form SHO monthly, and requires the SEC to publish the data aggregated by security through EDGAR. That aggregate would be a US Government work in the public domain, and it is the open source this dataset is waiting for.

It has been delayed twice. From SEC Release No. 34-104303 of 3 December 2025:

IT IS HEREBY ORDERED … that the Commission grants the temporary exemptive relief … from compliance with Rule 13f-2 and Form SHO reporting effective January 2, 2026, and ending January 2, 2028. As such, Form SHO reports for the January 2028 reporting period would be required to be filed within 14 calendar days after the end of January 2028.

First filings: mid-February 2028. Aggregated publication follows. A search of EDGAR full-text on 13 September 2026 returns zero filings on form type SHO, which agrees.

recipe/sources/form_sho.py is written and dormant. It writes nothing rather than writing placeholder rows that would later be mistaken for observations.

What is actually in here

File
metadata/source_licenses.yaml	the audit: terms URL, date read, clause, conclusion, per source
metadata/licensing_audit.parquet	the same thing as a table
metadata/coverage.parquet	every subset, its row count of zero, and the sentence explaining it
recipe/licensing.py	the gate, in code
recipe/sources/finra.py	the FINRA adapter, disabled from the public build
recipe/sources/form_sho.py	the SEC adapter, dormant until 2028
recipe/schema.py	every subset's schema, ready for the day there are rows

The subsets, and what each is waiting for

Every schema is written and every one is empty. Naming them is not decoration: when a source becomes publishable, the shape of the answer is already decided and the only thing that changes is whether rows are written.

Subset	Waiting for
short_interest	FINRA terms that permit this publication pattern
publication_calendar	the same — the schedule is FINRA's own website content
revisions	the same, and see Revisions below
sec_form_sho_positions	the SEC to begin publishing, February 2028
sec_form_sho_daily_activity	the same
instruments	securities to map; the table preserves the symbol a security had on a settlement date rather than rewriting history with today's ticker

The gate is code, not a promise

metadata/source_licenses.yaml is read by the build, not just by people. Every subset names the source it comes from, and a subset whose source says raw_redistribution_permitted: false cannot enter the published tree:

>>> from recipe.licensing import refuse_if_restricted
>>> refuse_if_restricted("short_interest")
RestrictedData: 'short_interest' comes from finra_equity_short_interest
(Financial Industry Regulatory Authority), whose terms as read on 2026-09-13 do
not permit publication here (status: restricted; see
https://developer.finra.org/specific-terms-equity-data).

There is no override flag. The only way to change the answer is to change the audit, which means re-reading the terms and writing down what they now say.

tests/test_licensing.py runs the whole build and then inspects what it wrote, so the claim "no restricted rows were published" is checked rather than asserted. The gate also scans the tree for files that merely look like a raw FINRA download, because a CSV saved during development would otherwise be uploaded without anything noticing.

Using the FINRA data yourself

FINRA's terms do permit your own non-commercial personal or professional use. The adapter is here for exactly that, and it will not run by accident:

export FINRA_CLIENT_ID=...        # your own credentials, having accepted the terms
export FINRA_CLIENT_SECRET=...
export SHORT_INTEREST_PRIVATE_DIR=/somewhere/outside/this/repository
python -m recipe.cli private --settlement-date 2026-08-31

It refuses to run without credentials of your own — there is no shared key here and none should be added. It refuses a private directory inside this repository, because that is the tree that gets published. It writes a LICENSE-RESTRICTED.txt beside its output, so a file found on a disk two years from now still says what may be done with it.

Collecting is not publishing. The restriction is on redistribution.

The PIT semantics, for when there is data

The whole reason to build this dataset rather than copy a snapshot is the lag, and the adapters already model it:

settlement_date	when the positions were held
firm_due_at	6:00 p.m. Eastern, two business days later — a reporting deadline, not a publication
knowledge_at	after the close on the publication date, about eight business days after settlement

Making the data visible to a backtest at the reporting deadline would hand it more than a week of positioning that nobody outside FINRA could see. tests/test_sources.py asserts the ordering.

Fields that will stay empty

short_float_pct, days_to_cover, short_interest_pct_float, borrow_fee, utilization and shares_available_to_short are not computed. Each needs a denominator — public float, shares outstanding, average daily volume from a licensed source — that this dataset does not have. The columns exist in the schema and stay null so that their absence is visible rather than silently filled with a number from somewhere unnamed.

days_to_cover is worth a specific note: FINRA publishes a figure of its own, and it is kept under its own name rather than merged into this column, because the two are computed against different volume windows.

Revisions

historical_revision_reconstruction = unavailable.

FINRA corrects prior cycles, but the source exposes only the latest corrected value. Where a vintage was not captured at the time, the pre-correction value cannot be recovered, and this dataset does not invent one. The revisions subset can only ever contain revisions observed prospectively.

What this dataset will never be

• FINRA daily short-sale volume used as a stand-in. Short interest is open positions outstanding at a reporting date; short-sale volume is the volume of short-marked executions during a day. They are different economic quantities and substituting one because it is easier to obtain would be a silent lie.
• The same data from a different door. No mirror, no vendor, no second FINRA interface, no relabelled source. A restriction routed around is still a restriction broken.

Where to get the data instead

• FINRA — register at the FINRA Developer Center, accept the terms, and use the Query API for your own non-commercial use. The adapter here does the rest.
• Your broker or data vendor — most redistribute short interest under a commercial licence they hold.
• The SEC, from 2028 — and when that happens, this dataset stops being empty.

The schedule re-reads the terms, not the data

There is no data to refresh, so the two Hugging Face Jobs do the only thing worth automating here: they fetch the terms pages this conclusion rests on and check whether they have moved.

Job	Cron (UTC)
us-short-interest-update	0 12 1 * *	monthly: re-read the terms, check the SEC date
us-short-interest-sweep	0 12 * * 1	weekly, so the February 2028 date is not missed by four weeks

FINRA's API terms print their own revision date, so a change is detected by what the page says about itself rather than by a digest that any redesign would break. When something moves, the run reports it and stops. It never decides on its own that something has become publishable — a machine noticing that a page changed is not a machine having read it.

Related

The other ZipLime point-in-time datasets, none of which hit this problem because their sources are US Government works: us-economic-events, commitments-of-traders, fomc-events, us-treasury-events.