Parsing 20,364 earnings press releases without guessing
When a US company reports earnings, the press release lands on SEC EDGAR within minutes, attached to an 8-K filing under Item 2.02. Veritas Data has parsed 20,364 of them. This is the pipeline, and the two rules that shape every line of it.
Rule one: nothing is inferred. A value the press release does not state is null, with a reason code. There is no "probably meant" in the dataset.
Rule two: every number gets checked against a second document the company also filed. The press release is written by the company for the market. The XBRL in the 10-Q or 10-K that follows weeks later is structured and audited. If they disagree, the response says so. We never pick a side.
The pipeline
- One EDGAR client for the whole system, capped at 8 requests per second with a contact email in the user agent, per SEC fair-access guidance. A lease in the code means a second process waits instead of adding load.
- The submissions feed surfaces new 8-Ks. Item 2.02 marks results announcements. The press release is almost always exhibit EX-99.1.
- Extraction is regex and document structure, not a language model. Tables first, then labeled sentences. GAAP before non-GAAP, with the label kept. Where extraction fails, it fails loudly: null, plus a reason.
- The release maps to a fiscal period from the company's own filing history, which is where fiscal-year shifts and transition periods try to ruin everything.
- When the 10-Q or 10-K arrives, we compare. EPS within one cent and revenue within 1.5% count as a match. Anything else flags the release, and the API shows both values.
The numbers
Across 20,364 parsed releases, measured November 19, 2026: EPS was readable 91.9% of the time and matched the XBRL 98.2% of the time. Revenue: 86.4% readable, 97.6% matched. The gap to 100% readable is real life: numbers inside images, scanned PDFs, prose-only releases, tables built from spaces.
Why not a language model
Failure mode. A regex that misses returns null. A model that misses returns a plausible wrong number. In a dataset whose whole point is that every value has a document behind it, plausible-wrong is the worst possible output. Determinism also makes the accuracy figures mean something: the same input parses the same way every time, so a published match rate is a property of the system, not of a sampling day.
What still breaks
In honesty: amended 8-Ks that restate a number, foreign private issuers reporting on 6-K instead of 8-K, companies with a 53-week year, and the occasional release where GAAP and headline figures swap places. Each one becomes a test case.
Everything above is queryable over REST and MCP, and every parsed value carries its source_url, so you can check our work.
Get a free key, 100 calls