Methodology

Models reason. Code calculates. Filings decide.

The principles behind Stoquity’s research: how it is produced, how it is challenged, how it is judged, and how the record is kept. Each task goes to whatever can be trusted with it.

Principles

Five commitments

Written down first
Every rating and forecast carries its date and the share price at the time. It is on the record before the outcome is known.
Falsifiable
Every rating states the conditions that would prove it wrong. Every forecast is a claim that a company filing will settle as true or false.
Independent
Several models from different vendors answer separately. None sees another’s answer. Stoquity takes no payment from the companies it covers.
Computed, not typed
Financial figures, ratios, returns and risk statistics are calculated by code from source data. Models reason about the figures; they never supply them.
Append-only
Nothing on a ledger is edited or removed. A mistake is corrected by a new dated entry beside the original.
Division of labor

Three kinds of work, three kinds of worker

Most failures of AI in finance come from asking a model to do something it should never be trusted with. Stoquity separates the work.

Models reason
They read a company’s record, weigh the case for and against, and commit to a judgment. This is what they are good at.
Code calculates
Every number on the site is derived by deterministic code from filings and market data, and can be traced to its source. A model cannot invent or alter one.
Filings decide
Whether a forecast came true is read from what the company later reports to its regulator. No model, and no person at Stoquity, marks the work.
Ratings

A view, and what would change it

Each company carries a rating on a four-step scale from Conviction to Cautious, with a stated level of conviction. A rating is never published alone. It comes with its break conditions: the specific developments that would turn it more positive or more cautious.

Each rating commits in advance to the horizons over which it will be judged. When a horizon arrives, the share’s performance is compared with its sector peers over the same period, and the result is recorded as a hit, a miss, or in between. Scores are written once and never back-filled.

Forecasts

The same questions for every company

Before each quarterly report, every covered company is asked the same kinds of question: about its revenue, its earnings, its cash generation and, where it reports one, its order backlog. Each question has a threshold set by rule from the company’s own reported history, so no one chooses an easy target.

Standard questions are what make the results mean something. They let one company be compared with another, one model with another, and this quarter with the last.

Models

Independent by construction

The questions go to frontier models from Anthropic, Google and Mistral. Each receives identical information and answers alone, as a probability with its reasoning. Disagreement is kept and shown, because where capable models split is where the real uncertainty lies.

Every answer records the exact model that gave it. The harness that runs the models is indifferent to which models they are, which is what allows an institution to seat its own.

Scoring

Judged on calibration, against a real bar

A probability is only useful if it is calibrated: things called 70% likely should happen about 70% of the time. Each model is scored with a standard measure that rewards exactly that, and punishes confident error more than honest doubt.

The bar is not zero. A model is compared with what history alone would have predicted for that kind of question, so it earns credit only for knowing something the base rates do not.

No forecasts after the fact. If a forecast was written when its answer could already have been known, it is withdrawn, marked as withdrawn, and never scored. 62 of 3,093 have been withdrawn so far. They remain visible.

The record

Kept so that it can be checked

Each company’s ledger is append-only. When a figure turns out to be wrong, a dated correction is added beside it and the original stays.

Every version of every record is fingerprinted, and each day’s set of fingerprints is independently timestamped outside Stoquity’s control. The date on any entry can therefore be verified without taking our word for it.

Coverage

What is covered, and how far back

1,090 US-listed companies have a ledger.

Part of the recordReaches back to
Quarterly financialsThe last eight quarters
Annual financialsThe last four fiscal years
Corporate events2017 for the typical company; 2004 at the earliest, 2022 for the most active filers
Share prices and risk statisticsMay 2024
Ratings and their scoresJune 2026
Three-model forecastsNovember 2026

Sources are public regulatory filings, licensed market data, and public datasets from government and international bodies, joined to the filings where they bear on a company.

Limits

What it does not do yet

  • Foreign filers. Companies that report under international standards do not yet have a full timeline or forecasts.
  • Banks, insurers and property trusts. Fewer forecasts apply, because some standard measures do not describe them.
  • Outside data. Deepest for aerospace and defense today.
  • A short track record. Ratings have been scored since June 2026, and the earliest three-model forecasts settle in late November 2026.
  • Scenario tools. Some scenario sizes are illustrative and are marked where they appear.
Going deeper

What is published, and what is not

This page sets out the principles and the rules we hold ourselves to. The detailed specification, including the question definitions, the scoring parameters and the controls around the models, is shared with institutional clients under a confidentiality agreement.

For institutions