assumption-drift

What this is

NVIDIA Corporation, Salesforce, Inc. and Target Corporation made numeric forward guidance statements in their own SEC filings, filed between January 2019 and September 2026. This dataset pairs each one with what the company later reported for that same metric and period, and records whether the company ever acknowledged the gap when the guidance was missed. It is a problem statement in data form, not a model and not a demonstration of one.

How a row is built

A row starts as a sentence or a table in an 8-K, 10-K or 10-Q filing that names a metric, a number or range, and a period. A later filing reporting the same metric for the same period supplies the outcome, when one exists. Every candidate row was drafted by a language model reading the filing text.

The resolution rubric

A row resolves mechanically from its own numbers, once the period has closed and an outcome exists:

• met: the reported value falls inside the guided range, endpoints included (a single point value is met within half a percent of itself)
• missed: the reported value falls outside the guided range, on the good side of it (better than guided) or the bad side (worse than guided). Which side is good depends on the metric. Higher is better for revenue, gross margin, operating margin, operating income, EPS, other income and expense, comparable sales and free cash flow. Lower is better for operating expenses and tax rate
• withdrawn: the company explicitly withdrew or suspended the guidance in a later filing, before the period closed
• unresolved: the period has not closed yet, or no later filing reports the metric

Acknowledged means a later 8-K, 10-K or 10-Q states the gap in so many words, naming which side of its guidance the result fell on: for example "below", "short of", "did not meet", "higher than expected", "exceeded" or "above the high end".

As of 2026-10-01, the release holds 611 rows: 103 met, 285 missed, 221 unresolved, 2 withdrawn. Of the misses, 221 were better than guided and 64 worse than guided; median days from the guidance to the filing that made it checkable was 92.0, and 97% of misses were never acknowledged in a later filing.

By company:

company	rows	met	missed	withdrawn	unresolved	never-acknowledged share
NVIDIA Corporation	212	58	79	0	75	100%
Salesforce, Inc.	317	18	168	0	131	99%
Target Corporation	82	27	38	2	15	79%

Provenance guarantee

Every source_url in this dataset returned a 200 from www.sec.gov or efts.sec.gov; no row was built from a URL that was not actually fetched. Every excerpt is the exact text taken from that filing. Its provenance hash is the sha256 of the extracted text, not of the raw page: SEC serves a per-request script tag inside the raw HTML that differs on every fetch of the same document, so only the extracted text is stable enough to hash and check again later. Every row was independently checked by a person against the cached filing before it was approved, and at least ten percent of the approved rows for each company were separately re-found on EDGAR by hand, their URL compared to the one the pipeline used.

Known limitations

• This covers three companies. It is a pilot, not a survey of the market.
• The pipeline does not claim to capture every guidance statement a company ever made; it captures the ones its patterns and its model caught.
• A row reported better than guided is recorded as missed, the same as one reported worse, because the guided range was wrong either way. Which side it fell on is reported separately, not folded into the status. Whether a side is good or bad is fixed once per metric, the same for every company: operating expenses and the tax rate are treated as lower-is-better, everything else as higher-is-better.
• A one-sided floor ("at least X") resolves as met on any reported value above the floor. There is no ceiling to measure against, so a floor's better-than-guided rows are not measured; only values below it can be missed.
• The search for an acknowledgement covers 8-K, 10-K and 10-Q filing text only. A company that only addressed a miss on an earnings call, and never wrote it into a filing, is not found: that text is not in EDGAR.
• An outcome that would come from a multi-column table whose header could not be confidently matched to a period is held back as unresolved rather than guessed at.

How to cite

If you use this dataset, please cite it as:

Navneet (2026). assumption-drift. https://huggingface.co/datasets/Nav772/assumption-drift. Accessed 2026-10-01.

Licence

The dataset is released under CC-BY-4.0. The code that built it is released separately under MIT.