Validation
Every figure on this page is a count over the record, never a probability. The thresholds were set in June 2026 and July 2026 with the whole history in view, so the backtest is in-sample; each feature’s live record starts on its card below, the earliest in July 2026. A result that new data can change is served, so this page moves with the record.
- What it reads
- Three gauges, each ranked within its own history, and the regime read from them by fixed rules.
- Version
- 0.3.2-m41 (the live engine); the point-in-time engines 0.5.1-fs-vintage, 0.5.1-ir-vintage and 0.5.1-mom-vintage.
- How it was tested
- Pre-registered: the criteria were written down, with kill conditions, before any measurement ran, in July 2026. The sample is the 4 NBER recessions the point-in-time timeline can be tested on (1990, 2001, 2008 and 2020), over a 45-year timeline — at least 85% first-published from Sep 2016; earlier years partly use today's revised data, read only once it had been published (how much).
- What it got right
- The Contraction label was entered within 1 to 5 months of every NBER peak (4 of 4); a placebo catches at most 2 of 4 by chance, and the adjudicated register holds 0 Contraction entries that were false alarms. The Sahm rule, under the same window, has fired 1 time since the timeline opened with no recession inside or after (2024); an OECD leading-indicator rule was earlier on 3 of 4 recessions at a cost of 18 false episodes, the frozen July 2026 count. Forward Stress's 12-month change fell 15 points or more 10 to 12 months before 4 of the 4 recessions its record reaches. The gauges’ rules were set in June–July 2026 and the stages’ in September 2026, with this history in view; the regime page’s live record starts September 2026 (2 months so far). Momentum has been below 30 for two months running 10 times since 1964, with 0 false alarms.
- What it got wrong
- The 1990 recession was the weakest catch: the label arrived 5 months after the peak, on the thinnest point-in-time coverage of the 4 (how much). Forward Stress's slide fired 5 times with no recession after it (1995, 1996, 1998, 2015 and 2022), and it read below 40 with no recession following 3 times (2022 and 2024). The credit-stress input understates market-led, sectoral and policy-masked credit episodes (2001, 2015–16 and 2020); that is documented, not repaired. The historical look-alikes failed the bar at 3 of 3 horizons, replayed over 477 months as first published — at least 85% first-published from Sep 2016; earlier years partly use today's revised data, read only once it had been published (how much), and the view was retired in September 2026. At the first quarterly sitting (October 1, 2026 to October 2, 2026), the trajectory calls for April 2026, May 2026 and June 2026 held 2 of 9; 2 went the opposite way, and the stress signal had 0 firings.
- Known limits
- The regime label recognises a recession; it does not lead one. Only the newest recession falls in the fully vintage-clean window, and the NBER dates recessions with a lag of around a year.
- This feature’s live record since
- July 2026 · 4 months · the four-stage regime since September 2026, with 6 stress episodes on its record from August 1981
- What it reads
- How many of the real-activity indicators are improving, deteriorating or abstaining each month, each judged against its own median band.
- Version
- 2.0.0-eb
- How it was tested
- In-sample, with the bars fixed before measurement: two tests over the census from February 1964 to October 2026 (753 months). The survey of record was measured in August 2026 and re-run on the record as it was known in September 2026.
- What it got right
- Test one, a persistent series that visits both extremes: it passes on every bar. Test two: of the 74 most-negative months among 633 eligible, 48 were followed by a recession month within six months (64.9% against a 16% rate across all eligible months) and 50 within twelve (67.6% against 21.6%); both counts clear the placebo bars.
- What it got wrong
- 24 of the 74 extreme months were not followed by a recession within twelve months; 4 of them sit clear of any recession in either direction (December 2002, January 2003, March 2003 and July 2022).
- Known limits
- The second test says the extremes cluster near recessions; it does not say that any one month predicts one. The census describes the present.
- This feature’s live record since
- August 2026 · 3 months
- What it reads
- Where the housing sector sits in its own cycle: voters in three timing stages name the stage, and a comparator says whether housing leads or lags the rest.
- Version
- 2.0.0-hc
- How it was tested
- Pre-registered: the roster's forecasting correlation, measured before the roster was chosen, in August 2026, and found low, so the census describes and does not forecast.
- What it got right
- Recession concordance, reported for the record and never a selection criterion: 82.5% of NBER recession months since March 1980 carry a downturn label (47 of 57). The comparator reproduces Economic Breadth without housing month for month (r = 0.983 over 659 months).
- What it got wrong
- 20.1% of the months outside a recession carry a downturn label too (101 of 503).
- Known limits
- No Housing Cycle output is described as leading, predicting or signalling. The record lists what followed as individual rows.
- This feature’s live record since
- September 2026 · 2 months
- What it reads
- How many of the oversold gauges sit at extremes against their own history, and what followed past washouts by the stage the economy was in.
- Version
- 0.1.0-capmon-A
- How it was tested
- Not validated against outcomes: this feature describes, it does not call. It is in no scorecard roster.
- What it got right
- Not scored. The feature makes no call, so there is nothing to get right or wrong; its record is a historical reference. The record: 82 washout episodes since August 1981; the S&P 500 was higher a year on after 58 of them, a median of 17.2%; in a slowdown, after 3 of 6, a median of −3.5%.
- What it got wrong
- Two stated limits, in the methodology: in November 2007 the Monitor fired while the economy read Slowdown, and the following year was deeply negative; the 1984 grind correction never reached Extreme Washout.
- Known limits
- The Monitor fired in a washout the economy read as a Slowdown, not a Contraction, before the deepest fall of its record; a slow grind can trip the count without ever reaching an extreme washout.
- This feature’s live record since
- July 2026 · 4 months
- What it reads
- How expensive the market is against its own history, on four lenses and two bases.
- Version
- unversioned
- How it was tested
- In-sample: every eligible month of each lens's record is part of exactly one period, measured from the month its first reading was published; "resolved badly" defined before anything was computed. Measured September 2026.
- What it got right
- At three years the CAPE record tilts: expensive periods came in below the rest of the record's median 16 times out of 20 (a median of 21.4% against 41.2%), and on the earnings-yield lens 23 out of 30. At one year the earnings-yield lens came in below the rest 24 times out of 33.
- What it got wrong
- At one year the CAPE record says nothing a warning could use: expensive periods returned a median of 15.2% against 9.6% for every other period, and were negative 6 times in 21 against 16 in 47. 2 of the lenses (buffett indicator and household equity allocation) have fewer than 30 periods and cannot be sorted.
- Known limits
- The lenses with short records cannot be sorted into a finding, and the record of the longest lens largely predates the engines it was tested against.
- This feature’s live record since
- July 2026 · 4 months
- What it reads
- Corporate profits as a share of the economy, measured against their own trend on a five-level ladder.
- Version
- unversioned
- How it was tested
- Pre-registered: the two claims written down in 1991's terms before the test ran. The register holds 56 entries; the what-followed census was measured October 2026 over 1963 to 2025.
- What it got right
- Early recoveries are under-reported: VALIDATED, on 3 qualifying recoveries with true-vintage coverage, every one first reported weaker than it proved. Of 55 resolved periods, 6 lost money over the next year (11%, against a base rate of 22%).
- What it got wrong
- Margin mean reversion: WATCHABLE, on 4 episodes, with 0 reversions among the episodes since the century turned, at any horizon. Of 18 conditions tested as marks of the losing periods, 12 pointed the wrong way and 1 separated.
- Known limits
- A handful of episodes can neither validate nor kill a claim; the record grows as new episodes complete, and the annual revision moves the register.
- This feature’s live record since
- July 2026 · 4 months
- What it reads
- The past months that look like today, assembled from the other features’ maintained records.
- Version
- unversioned
- How it was tested
- Not validated against outcomes: this feature describes, it does not call. There are no Confluence tables and no Confluence jobs; the page assembles at load from the other features' records.
- What it got right
- Not scored. The feature makes no call, so there is nothing to get right or wrong; its record is a historical reference. The register: 30 closed episodes over 68 matching months, settled through October 2026.
- What it got wrong
- Not scored: the register's open forward columns are resolved at the quarterly sitting, and the page makes no claim about them.
- Known limits
- A match is a description of resemblance, never a signal; what followed the matches is counted in the register, and the open forward columns wait for the quarterly sitting.
- This feature’s live record since
- July 2026 · 4 months
- What it reads
- Every inversion of the three-month to ten-year Treasury curve and whether a recession followed, as tested.
- Version
- unversioned
- How it was tested
- Pre-registered: one rule and a five-part test fixed before measurement. Each inversion is scored on whether an NBER peak followed within 18 months, once 12 months have passed after its window closed; measured September 2026.
- What it got right
- The curve scored 7 of 9 against a base rate of 16.5%, 4.73 times as often; 500 random placements scored a median of 1. The two-year curve, tested the same way, scored 8 of 11. The curve came before the site's own stress signal 2 times of 7 and before the stage 4 times.
- What it got wrong
- 2 inversions were followed by no recession (2006 and 2022). The two-year curve missed the recession that began in Feb 2020 and had 3 false alarms. The 8 cross-asset relationships this page once followed were tested by the same rule, and 7 did not hold.
- Known limits
- An inversion is scored only after the window and the announcement lag have passed, so the newest inversions read as not yet declared, not as misses.
- This feature’s live record since
- September 2026 · 2 months
- What it reads
- How far first prints miss the settled record, series by series, from the archived vintages.
- Version
- spine 2
- How it was tested
- Pre-registered: three censuses with their bars fixed before the vintage walk ran; measured September 2026 over 8 dossiers.
- What it got right
- First prints lean with the cycle: on payrolls, the lean between the expansion and contraction stages holds at p < 0.000001.
- What it got wrong
- Revisions are not larger near turns: 7 of 8 series failed that bar over 5 recessions (gdp passed, on the turns its record reaches). An episode of revisions does not warn of a turn: 4 of 4 candidate cells failed, the best of them 8 hits in 63 firings over 10 turns.
- Known limits
- The dossiers cover the series with a usable vintage archive, and the floors are honest: a naive recomputation gives different figures.
- This feature’s live record since
- August 2026 · 3 months
What was tried and dropped
Ideas are tested before they stay on the site, and the failures are published with the successes. Each line names the date the record closed on it and the record itself.
- The historical look-alikes (the analog engine) — September 23, 2026. Replayed month by month on data as first published, the closest past matches came no closer to what followed than random picks at any horizon, against a bar fixed before the run. The view left the page. 3 of 3 failed. The record.
- An episode of downward revisions as an early warning — September 22, 2026. Two candidate signals at two settings were tested against a five-part bar fixed in advance; every cell failed, and the best alarm fired barely more often before a turn than at any other time. 4 of 4 failed. The record.
- The eight-relationship market monitor — September 25, 2026. Each relationship was tested against the outcome its own page claimed, by one rule fixed before measurement; none held. The yield curve, tested the same way, did, and has its own page. 7 of 8 failed. The record.
- The eleven-label regime engine — September 19, 2026. A census of its own record found labels that flickered on a month's noise and pairs of labels the following year could not tell apart. Four stages on one gauge, censused before they were ratified, replaced it. The record.
- Each indicator's own reading as a warning — October 4, 2026. Every published indicator's reading was tested, pre-registered, as a warning of a recession or a market fall within a year; no test cleared the bar, and the study is on the methodology page. 612 tests on 112 indicators; 0 passed. The record.
The scorecard and the changelog
Every quarter, what the site published is scored against what happened, as counts, and recorded. The first sitting, for the third quarter of 2026, was held on October 1, 2026 to October 2, 2026; the next is due on December 30, 2026. How a sitting grades the calls.
What changed on the site, batch by batch, is on the changelog. Changelog.