Methodology

Every feature on this site follows the same rules. Each feature’s own methodology page documents how they apply in detail — these are the commitments they share.

Historical record, never a forecast. Everything on this site describes what is and what has been — the current state of the economy, the full history of past episodes, where today’s readings sit against historical ranges. Nothing on this site is the site’s own forecast; two heatmap cells relay others’ — the Atlanta Fed’s GDPNow nowcast and the Chauvet–Piger smoothed recession probability (via FRED). A reading at a level historically associated with stress means exactly that, and no more.

Not a trading signal. No page tells you when to buy or sell. The site shows the macro backdrop against which investment decisions are made; the decisions are yours.

Point-in-time honesty. Where the site says “this is what was knowable at the time,” historical views use the data as originally published, not as later revised — wherever an archived first print exists. Where none does, the older record is partly today's data, and the share is measured. The State of the Economy’s Time Machine and point-in-time timeline and the Regimes page are built on this basis — and where first-print and revised history disagree, the site says so. On a day the site recorded its live page, the Time Machine shows that record instead. On a day the site has no record of its live page, the heatmap’s Time Machine reads what had been published by the picked day — first prints where an archive holds them, today’s value otherwise. On a recorded day it shows the heatmap as it was.

Counts, never probabilities. Where the site reports what followed past episodes, it reports counts of named, checkable events, and the share they make of the record — never odds on the next one. Small historical samples don’t support probability language, so the site won’t manufacture it.

The 7,399 claims on this site that have an automatic test were last re-checked on November 8, 2026: all 7,399 held. Since October 5, 2026, 94 claims have moved at a re-check: 59 corrected, 35 moved with new data, 0 open. Each is listed on the Corrections page.

History as it was known. When the site looks back, it uses each figure as it was first published, on the day it came out. It does not use the cleaner numbers that came later, or hindsight. Before 2016, not every first-published figure was archived, so part of that older history uses today’s revised figures (how much) — and two small, measured exceptions in the Forward Stress record are set out there.

Every number is worked out or checked. Each number on the site is either calculated from the data every time the page is built, or written down with a test that checks it. A check stops anyone adding a number that is neither.

How episodes are counted. When a page counts how many times something happened, such as periods of extreme fear or crossings of a line, one rule separates one episode from the next. For a daily or weekly series, readings less than three months apart count as one episode. For a monthly or quarterly series, an episode is consecutive qualifying months or quarters, and one off ends it. A page says so only where its count uses a different rule: a line that must be crossed back before a new episode counts, a stated minimum length, or a gap of its own. “Under way or began within a year” counts a recession whose first month is the episode’s month or falls in the twelve months after it, twelve months exactly included. A month counts as inside a recession from the month after the peak through the trough, as in FRED’s recession series.

Re-checked every month. Once a month the site re-runs every test it can run on its own. It runs them again straight after big events, such as a recession being officially dated or the government’s yearly revision of its figures. Anything that has changed is flagged for a person to fix.

Its calls are kept and graded. When the site flags a change in the economy, such as stress building or a trend turning, that call is saved on the day it appears. Every quarter, the saved calls are scored against what actually happened (how).

Most ideas fail, and the failures are published. Ideas are tested before they go on the site, and most don’t make it. Of 8 classic market pairs tried, such as copper against gold, 1 held (the yield curve) and 7 did not (the record). Matching today with its closest look-alikes in history failed at all 3 horizons tried (the test). An episode of downward revisions to the data, tested as an early warning of recessions, failed all 4 ways it was tried (the findings). The readings of 112 indicators were also tested as warnings of recessions and of market falls: 612 tests, and none passed (the study).

A description, never advice. The site describes the economy and its record. It never tells anyone what to do with their money. It makes no forecasts of its own; the two forecasts it shows, the Atlanta Fed’s GDPNow and a recession probability, are other people’s, and say so.

Could an indicator page’s reading warn of what tends to follow? In October 2026 the site tested that across its indicator pages, with the rules fixed before any result was seen. The answer was no: of 612 tests on 112 indicators, none passed.

The two events. Each test asked whether a reading came before one of two events. One is the economy’s turn into a recession within 12 months: a business-cycle peak, as dated by the National Bureau of Economic Research, the accepted record of US recessions. Months already inside a recession were left out. The other is a stock-market fall: the S&P 500, with dividends reinvested, closing at least 10% below that month’s level at a month-end within the next 12 months.

The readings, as they were known. Each page’s figure was read as it stood at each month’s end: the latest figure published by then, at its value that day where an archive of first-published figures holds it, and at today’s revised value where none does.

The ideas. Past months like the reading at the time of the study were picked out three ways:

• in the same colour band of the page’s scale;
• a change over 6 months in the same fifth of all such changes;
• in the same band for at least as long (tried only where that period had lasted 3 months or more).

Each was tested against both events: up to 6 tests a page. A signal counted once each time the reading entered the condition, however long it stayed.

The bar. A test passed only by clearing all 5 parts of the test the site’s yield-curve record is held to:

1. Enough history: at least 5 signals from at least 4 episodes (signals under 12 months apart are one episode).
2. Better than usual: the event followed at least 1.5 times as often as after an average month.
3. No more false alarms than hits.
4. Not resting on one episode: parts one to three still hold with any one episode left out.
5. Better than chance: more hits than random months scored in at least 95 of every 100 of 2,000 tries, and more than that when a page had more tests.

A near miss, also set in advance, was a test that failed just one part, by a small margin.

What held: nothing. Of the 115 indicator pages, 112 were tested, 612 tests in all. None passed all 5 parts, and none was a near miss. The other 3 had fewer than 36 monthly readings and were not tested: ISM Manufacturing PMI® (2 readings), ISM Services PMI® (2) and Existing Home Sales (17).

How far short. On every one of the 112 pages, even the best test failed “not resting on one episode”: leaving out a single past episode undid it. The best test on each page, by how many parts it failed:

The closest. U-6 Unemployment Rate, in the same colour band: 8 of its 16 signals were followed by a stock-market fall, 2.57 times as often as after an average month. It failed only “not resting on one episode”, by more than the near-miss margin. The table lists every page whose best test failed no more than two parts.

What this means. No tested reading, on its own, had a record strong enough to be called a warning by this test, so no sentence about what usually follows was added to any page. This is a record of a test, not a forecast: it does not say that a recession or a market fall is more or less likely now.

Run again. The study is re-run every year after the government’s yearly revision of its figures each September, and whenever a recession’s start or end is officially dated, with the same code and the same test.