Point in time
Each observation records what was known on the day it was made. Snapshots are written once and never rewritten, so a backtest built on them cannot quietly inherit information from the future.
Research data for public equities
Atomic Research publishes individual, point-in-time research observations drawn from SEC filings and government data. Each one is a single −1, 0, or +1, carrying the evidence behind it. Nothing is aggregated, nothing is ranked, and nothing is quietly revised later.
The unit of data
Every record answers one narrow question about one company on one date, and shows you where the answer came from. What you build on top of that is your decision, not ours.
{
"symbol": "MSFT",
"signal_id": "cash-flow.fcf-level-and-margin",
"date": "2026-08-18",
"state": "positive",
"value": 1,
"definition_version": 1,
"maturity": "deterministic",
"observed_at": "2026-08-18T06:12:04Z",
"evidence_date": "2026-06-30",
"source_method": "sec-companyfacts",
"sources": [
{ "kind": "sec-filing", "reference": "0000789019-26-000023" }
]
}
Each observation records what was known on the day it was made. Snapshots are written once and never rewritten, so a backtest built on them cannot quietly inherit information from the future.
missing, failed, and
inapplicable are distinct, explicit states. A signal we
could not observe is never silently converted into a negative one.
Every observation declares how it was derived, from arithmetic on a filing through to a model’s reading of a disclosure. You decide what deserves your trust, rather than discovering the limits afterwards.
Scope
The following are omissions by design rather than features still to come. The product is evidence. Deciding what it means is your job.
Provenance
Some observations are computed directly from filings and are fully reproducible. Others are language-model readings of material that resists arithmetic, such as competitive position or the credibility of management’s own claims. Both halves matter, and the second is the harder one to obtain.
Every observation says which kind it is. Model-derived readings can be wrong, so each one carries its maturity and the reasoning behind it, and you can filter on both. The underlying records are public.
The API is in active development. If this is data you would use, write and tell me what you would use it for. That genuinely shapes what gets built first.
help@atomicresearch.markets