How accurate is our outlook, really?

Definitions

Every projection is a directional call (up / down / flat) plus a range. Once the next filing lands, we grade it one of three ways:

• hit the actual number landed inside the projected range.
• partial right direction, but outside the range.
• wrong call we called the wrong direction.

This grades our forecast, not the company — a "wrong call" can still be great news for the business (e.g. we projected a decline and the actual number grew).

Inputs & exclusions

A projection is not the company's own guidance, and it isn't freely made up: the range comes from a deterministic statistical model — the empirical spread of what actually happened historically to companies whose metric was trending the same way (up / down / flat) as this one's most recent year. We do not use analyst estimates, guidance, or anything the company itself forecasts.

Net income's range runs noticeably wider than the other three metrics below. That's structural, not a modeling gap: net income is the GAAP line most exposed to one-time items (impairments, tax adjustments, asset sales) and to companies sitting near breakeven, both of which swing the percentage sharply. Treat the direction as the signal; the width is inherent to this one metric.

Backtest vs. live

The stats above are an out-of-sample historical backtest: each historical filing is graded against a range computed only from data available before that filing, the same way a live call would be — but run retrospectively across SEC history, so the record has real sample size from day one instead of starting empty. It is not the same claim as "these were live predictions we published in advance."

We also keep a separate, much smaller count of genuinely live calls — graded only after they were made and a real filing settled them going forward. It's shown below on its own, not folded into the backtest number above, specifically because it's too small yet to draw conclusions from.

Since public launch

Limitations

• The backtest is retrospective and out-of-sample, but it is not the same as a track record of calls published in advance — see "Backtest vs. live" above.
• Ranges are fitted from the historical spread across many companies, not personalized to any one filer's own volatility.
• This grades a directional call and a range, not investment performance — a "hit" is not a recommendation, and a "wrong call" is not investment advice either way.
• Figures on this page are computed live from the same database the app reads and will shift over time as more filings are graded.

Does a red flag predict an SEC enforcement action?

Our forensic score isn't only about fiscal projections — it also flags financial-statement red flags (leverage, accruals, liquidity, and three more deterministic checks). We wanted a harder test than grading our own forecasts: does that score actually run higher for companies the SEC later charged with accounting fraud than for the market at large?

We cross-referenced every SEC Accounting and Auditing Enforcement Release (AAER) since 2010 — the earliest year XBRL data exists to check against — matched each one to a specific public company by name, and reconstructed what our forensic score would have said the last time that company reported before the SEC acted, using only the data that was actually knowable at the time. The market-wide baseline runs the identical five checks across every US filer, same years, so "elevated" and "high" mean exactly the same thing on both sides of the comparison.

What we found

Loading…

Inputs & exclusions

An AAER is matched to a CIK only by an exact, deterministic normalized-name match — the same matching primitive the app uses elsewhere for 13F holdings. A name that collides across more than one real company (a parent and its own subsidiary, or two unrelated companies sharing a dual-listed name) is left unmatched rather than guessed. Audit firms, individuals, and private advisers named alongside a company in the same release are excluded entirely — only the company itself is scored.

A company with too little XBRL data to run any of the five checks is counted as its own bucket, not silently dropped and not counted as a miss — both here and in the market-wide baseline, so neither side is inflated or deflated by how much data happened to exist.

Limitations

• Scope is 2010 onward only — no XBRL data exists before SEC's mandate, so an earlier enforcement action could never be scored either way.
• Name-matching is exact and conservative by design (see "Inputs & exclusions"), which means some real matches are left out entirely rather than risked as a guess.
• This checks whether our score's elevated/high threshold discriminates enforcement companies from the market at large — it does not evaluate the score's five underlying checks individually, and a company scoring low was not thereby cleared of anything.
• The SEC brings enforcement actions years after the underlying conduct, sometimes a decade or more — a red flag from the last filing before an action is a proxy for what was visible then, not a claim about exactly when anything actually happened.