HighGravityDataMethodology
The source rule
A value appears on the site only if it can be located, word-for-word, in the filing body. When it can, we store its exact character offset so “view in filing” takes you straight to it. A value we cannot locate is listed as not located in filing with its cause — never shown as guidance. The rule holds at the data boundary: an item without a filing span is not exported at all. This is the whole product — a number you can check, not a summary you have to trust.
What we publish
For each company, every forward-looking guidance item stated in its quarterly earnings-release exhibits, quarter by quarter: the metric, the value and period in the company's own words, how we classify it, and a link to the exact span in the source filing.
How guidance is extracted
Within that rule, extraction is model-based: a language model reads each earnings exhibit and pulls out every forward-looking guidance item — a sales range, an earnings target, a dividend, a buyback, a multi-year goal — as metric, value, and period, copied from the filing rather than paraphrased. We then check the model's output against a careful human read of the same filings. It is not a hand transcription of every filing, and it is not fully automatic end to end.
How guidance is classified
Each item is classified by what kind of measure it is. A company-defined measure is one the company defines itself (a non-GAAP or operating figure), checkable only against the company's own reconciliation or later results. A GAAP-tied item maps to a standard audited financial line, directly or by a simple calculation. A declared action — a dividend or a buyback — is something the company says it will do, resolvable against the later filing. We record and classify what the company stated; we do not judge whether the company will meet it.
What the coverage figures mean
We check the model's extraction against a careful human read. On a probe of seven companies (n = 187 human-annotated guidance items), the model caught 91% of them, ranging from 82% to 100% per company — measured against a single human read, so it reflects one reader's judgment, not a consensus. We expect to miss some items; if you spot one we missed or got wrong, report it and we add it and log the correction.
We also check how we describe the guidance we surface — the change label on each item and the fiscal period we assign — against the filings. We do not publish an automated figure for this: an automated check would share the pipeline’s own extraction, so it cannot catch the pipeline’s errors. The result of an independent model review, n = 20 each:
- Change labels: 19 of 20 consistent with the filings.
- Target periods: 17 of 20.
Items we could not locate in the filing are surfaced and counted as “not located in filing” on the company page, never silently dropped. Where a whole sector or filing format resists machine reading, we say so below rather than imply full coverage.
Updated after each earnings release, typically within a few days; sampled human review follows. We process new earnings releases as they are filed; a person reviews a sample after each refresh. A new or reworded metric publishes and is queued for a merge decision — the worst case is one series shown as two until we merge them. Only a wrong exhibit, an image-only filing, or a filing we could not read (a format break) is held back before it appears.
Where coverage is thin
- Guidance stated without a number. Wording like “roughly flat” or “low-single-digit growth,” with no figure, is the hardest to catch and the main reason per-company coverage bottoms out around 82% rather than 100%.
- Metals & mining. About 30% of that sector's exhibits yield no machine-readable guidance — idiosyncratic table formats, or none — the highest rate of any sector.
- Image-only filings. Some issuers file the earnings release as page images; that guidance is not machine-readable for anyone, so those pages show little by disclosure format, not by omission.
- Wrong-file residual. A small number of exhibits are not the earnings release at all (mostly image-only, plus a shareholder-letter edge case); their guidance is not represented.
What remains proprietary
We publish the guidance, its classification, its filing link, and these coverage figures. The extraction models and prompts, the vocabulary that maps a metric to how we classify it, and our internal thresholds and checks remain proprietary.
Disclosure
HighGravityData records what a company's filing says and links it to the filing. It does not provide investment advice. An item shown as “not located in filing” is a description of what we could match, not a judgment about the company.
The words in quotation marks are the company's; the classification and the filing links are HighGravityData's. Every coverage figure on this site is reported with the number of companies it rests on.