Economic observatory
Model validation
Last updated
What the evidence can support, what has been tested, and what remains necessary before a forecast can be published.
Headline probabilities remain withheld. The current records are research diagnostics, and no qualifying regional probability is linked to this snapshot. A historical screen using revised data does not recreate information available in real time.
Validation at a glance
These cells describe the model linked to the current evidence. Separate historical studies remain available below.
| Check | United States | Euro area | United Kingdom |
|---|---|---|---|
| Target and horizon | RecordedA recession begins in one of the next 12 complete calendar months after the forecast-origin month, conditional on not already being in a recession. | RecordedA euro-area recession begins in one of the next four calendar quarters after the forecast origin, conditional on not already being in recession, using CEPR peak-excluded quarterly dating. The onset is identified only to the quarter. | RecordedA recession begins in one of the next 12 complete calendar months after the forecast-origin month, conditional on not already being in a recession. |
| Recession episodes tested | 3 onsets366 months; repeated forecasts are not independent episodes. | 0 onsets18 quarters; repeated forecasts are not independent episodes. | 0 onsets0 months; repeated forecasts are not independent episodes. |
| Historical performance screen | Not passedSince 2010: recent performance fails the descriptive base-rate/calibration screen. | Not passedFull results and subperiods are below. | Not passedFull results and subperiods are below. |
| Real-time information test | Not passedUses only releases and revisions available at each forecast date. | Not passedUses only releases and revisions available at each forecast date. | Not passedUses only releases and revisions available at each forecast date. |
| Independent holdout | Not passedA separate test of the fixed procedure. | Not passedA separate test of the fixed procedure. | Not passedA separate test of the fixed procedure. |
| Publication approval | WithheldResearch results do not grant approval. | WithheldResearch results do not grant approval. | WithheldResearch results do not grant approval. |
How activity, warning signals and evidence quality differ
Three separate questions
What is happening now?
Employment, income, spending and production describe the breadth of current economic weakness. This evidence assessment is not an official recession declaration or a validated numerical nowcast.
What might happen next?
A future-onset forecast has a specified region, horizon and current-state condition. It needs its own probability validation; a concern index cannot be converted into odds.
How dependable is the evidence?
Coverage, source freshness, historical availability and model validation are shown separately. More charts, more frequent checks and a complete AI brief do not by themselves establish forecast accuracy.
Corrected historical test
The new study trains each model on the histories available when it would have been fitted, then uses the information available at each later prediction date. This recovers recession episodes that the first replay consumed entirely in training.
3 of 3 candidates fail the declared research screens. The corrected test is more informative, but does not qualify a current probability. Recent-period performance and calibration remain obstacles.
| Candidate | Test dates / origins | Recession onsets | Brier error / baseline | Onsets detected | False-alarm episodes |
|---|---|---|---|---|---|
| Yield curve | 1997-01-31 to 2025-09-30317 forecast origins | 3 | 0.0943 / 0.1046 | 3 of 3 | 6 of 11 alert episodes |
| Labour and output | 1984-01-31 to 2025-09-30465 forecast origins | 4 | 0.108 / 0.1059 | 1 of 4 | 10 of 11 alert episodes |
| Small macro model | 1997-01-31 to 2025-09-30317 forecast origins | 3 | 0.0999 / 0.1046 | 3 of 3 | 8 of 13 alert episodes |
Lower Brier error is better. Repeated monthly forecasts are not independent recessions. Alert episodes group consecutive warnings; a false-alarm episode contains no matching recession event under the declared target. Detection uses the study’s fixed alert threshold and does not prove calibration.
Recent performance, warning episodes and limitations
| Candidate / period | Origins | Model error | Baseline error | False-alarm episodes |
|---|---|---|---|---|
| Yield curve · Since 2010 | 187 | 0.0839 | 0.0738 | 2 |
| Yield curve · Since 2020 | 67 | 0.1429 | 0.0472 | 2 |
| Labour and output · Since 2010 | 187 | 0.0866 | 0.0765 | 2 |
| Labour and output · Since 2020 | 67 | 0.0884 | 0.0503 | 2 |
| Small macro model · Since 2010 | 187 | 0.0938 | 0.0738 | 5 |
| Small macro model · Since 2020 | 67 | 0.1705 | 0.0472 | 5 |
These historical outcomes have already been examined during research. This is not an untouched holdout or a record of forecasts actually issued in the past. Training values can include revisions known at fitting, while prediction values reflect the information then available. That difference and delayed recession announcements remain material limitations.
Yield curve
- A recent-period candidate score is worse than its matched training-base-rate benchmark.
- A calibration window exceeds the predeclared descriptive error threshold.
Labour and output
- Full-sample Brier improvement below predeclared descriptive threshold.
- A recent-period candidate score is worse than its matched training-base-rate benchmark.
- A calibration window exceeds the predeclared descriptive error threshold.
- Recession-onset detection falls below the predeclared alert threshold criterion.
Small macro model
- Full-sample Brier improvement below predeclared descriptive threshold.
- A recent-period candidate score is worse than its matched training-base-rate benchmark.
- A calibration window exceeds the predeclared descriptive error threshold.
Study: us-fit-date-vintages-2026-10-27-v2; generated 27 Oct 2026, 00:53 UTC. No production approval is created by this study.
Protocol SHA-256: cd74847f4d687de7e39253c38285cb2d2f8e2e8a29ed75cfc61ac31edd5da12e.
Automated reproduction record
A separate implementation reproduced 1,099 historical forecasts and 9 score windows, with a maximum probability difference of 3.72e-9. This checks the calculations; it is not human review or publication approval. Recorded 27 Oct 2026, 00:57 UTC.
Reviewed result SHA-256: d10ef172b66cf247de98cbb1b193b272970d13c0f8922f7f38997f6050a7ab05. Review protocol SHA-256: cd74847f4d687de7e39253c38285cb2d2f8e2e8a29ed75cfc61ac31edd5da12e.
Current regional readiness
The following records describe the model saved with the live snapshot. The corrected historical study above and the regional archives below are separate research records. Their existence does not retroactively validate the legacy model or change its recorded input metadata.
| Region | Publication state | Evaluation sample | Brier error / base rate |
|---|---|---|---|
| United States | Research onlyvalidation-failed | 366 months3 recession onsets; overlapping forecast windows are not independent episodes. | 0.0905 / 0.0997 |
| Euro area | Research onlyinsufficient-data | 18 quarters0 recession onsets; overlapping forecast windows are not independent episodes. | 0.1183 / 0.0194 |
| United Kingdom | Research onlyinsufficient-data | 0 months0 recession onsets; overlapping forecast windows are not independent episodes. | Unavailable / Unavailable |
Brier error measures squared probability error: lower is better. A long test period containing no recession cannot establish recession-detection performance.
United States: research only · model detail
A recession begins in one of the next 12 complete calendar months after the forecast-origin month, conditional on not already being in a recession.
Recorded target window: 2026-12-01 to 2027-11-30.
Current-state condition: not confirmed. Dating convention: NBER monthly, peak excluded.
- The experimental model did not pass the historical research checks. Since 2010: recent performance fails the descriptive base-rate/calibration screen. Since 2020: recent performance fails the descriptive base-rate/calibration screen. Its numerical probability is withheld.
- Since 2010: recent performance fails the descriptive base-rate/calibration screen.
- Since 2020: recent performance fails the descriptive base-rate/calibration screen.
- The complete historical information set has not passed as-of release and revision checks.
- Independent evaluation of the frozen procedure remains incomplete.
0 of 45359 model-input observations carry availability dates; 0 carry vintage identifiers. These counts alone do not prove validation.
Euro area: research only · model detail
A euro-area recession begins in one of the next four calendar quarters after the forecast origin, conditional on not already being in recession, using CEPR peak-excluded quarterly dating. The onset is identified only to the quarter.
Recorded target window: 2026-10-01 to 2027-09-30.
Current-state condition: not confirmed. Dating convention: CEPR quarterly, peak excluded.
- The expanding-window check has 18 usable test quarters and 0 recession onsets. At least 20 quarters and one onset are required.
- The evaluation contains no recession onset, so it cannot establish recession-detection performance.
- The complete historical information set has not passed as-of release and revision checks.
- Independent evaluation of the frozen procedure remains incomplete.
0 of 17415 model-input observations carry availability dates; 0 carry vintage identifiers. These counts alone do not prove validation.
United Kingdom: research only · model detail
A recession begins in one of the next 12 complete calendar months after the forecast-origin month, conditional on not already being in a recession.
Current-state condition: not confirmed. Dating convention: No verified UK dating target.
- A verified UK recession-dating history is required before training a probability model.
- This live model has no verified historical recession labels for its declared target. A separate research target cannot be substituted silently.
- The evaluation contains no recession onset, so it cannot establish recession-detection performance.
- The complete historical information set has not passed as-of release and revision checks.
- Independent evaluation of the frozen procedure remains incomplete.
0 of 1327 model-input observations carry availability dates; 0 carry vintage identifiers. These counts alone do not prove validation.
Regional historical archives now integrated
The euro-area and UK research datasets now retain the underlying source files, checksums and dated revisions. These archives improve the historical evidence available for testing; they are separate from live monitoring feeds and approved forecasts.
| Series | Non-missing observations | Vintages | Recorded availability range | Economic periods |
|---|---|---|---|---|
| UK real GDPUK.BOE.ABMI | 26,876 | 163 | 1976-03-31 to 2016-09-30 | 1955-01-01 to 2016-04-01 |
| Euro-area real GDPEA.ECB.RTD.GDP | 17,624 | 254 | 2001-02-01 to 2026-09-11 | 1991-01-01 to 2026-04-01 |
| Euro-area employmentEA.ECB.RTD.EMPLOYMENT | 15,048 | 166 | 2001-04-01 to 2015-01-01 | 1980-01-01 to 2014-04-01 |
UK: the preserved Bank of England/ESCoE GDP archive covers vintages through September 2016. The ONS continuation after 2016 is still missing. Its quarterly two-decline confirmation target is not a broad monthly recession chronology.
Euro area: GDP vintages now extend through 2026, while the employment archive ends earlier. Changing euro-area membership and accounting definitions require within-vintage calculations; levels must not be spliced across definitions.
Source provenance and remaining limitations
- ecb-quarterly.zip ↗ SHA-256: f211651978b852aabbba577b00f150d488e8a3b35e7911cf727cee82cd83aa72
- uk-real-long.xlsx ↗ SHA-256: cba629f14c6ad3acb41b98e59492edfa8772f7c00664aca5ebd31e44c9ca7cd9
- ecb-gdp-history.csv ↗ SHA-256: 411003497ba0f842b9bb432c463fbbe22b13ec295be671b1bbe5e5a399940af8
- Research archives are not production feeds or a qualified forecast.
- UK preserved archive ends September 2016; an ONS continuation has not been acquired. IMF vintages remain a separate source and definition.
- Euro-area archive changes composition, base year and accounting standard. Compute growth within a single vintage; do not splice native levels.
- ECB API series metadata may describe the current sibling retrospectively; dates identify research-database freezes, not exact individual-source publication times.
- CEPR labels are only known at announced turning points. The unclosed post-2020 interval is not labelled expansion.
- UK labels are quarterly technical two-decline confirmation states, not a broad monthly recession chronology.
Archive recorded 27 Oct 2026, 00:56 UTC. Conservative availability boundaries may follow the labelled archive month; they are not invented exact publisher release times.
Euro-area and UK quarterly tests
These separate studies test whether the GDP histories available at fitting could warn about a specified event during the following four complete quarters. They do not inherit approval from a US model.
Euro area
The event is a recession starting in the quarter after a CEPR peak, conditional on expansion under the historical chronology. Unconfirmed periods after the last announced turning point are not invented expansion labels.
United Kingdom
The event is the first quarter confirming two consecutive falls in real GDP, conditional on that technical state not already occurring. This is distinct from a broad monthly recession definition.
| Region / candidate | Evaluated origins / dates | Brier error / baseline | Events detected | False-alarm episodes |
|---|---|---|---|---|
| Euro areaQuarterly GDP growth | 13 quarterly origins2015-03-31 to 2019-06-30 | 0.0724 / 0.0726 | 0 of 1 | 0 of 0 alert episodes |
| Euro areaGDP growth and momentum | 13 quarterly origins2015-03-31 to 2019-06-30 | 0.0741 / 0.0731 | 0 of 1 | 0 of 0 alert episodes |
| United KingdomQuarterly GDP growth | 130 quarterly origins1979-03-31 to 2015-06-30 | 0.1138 / 0.1119 | 0 of 4 | 2 of 2 alert episodes |
| United KingdomGDP growth and momentum | 129 quarterly origins1979-03-31 to 2015-06-30 | 0.1166 / 0.112 | 0 of 4 | 2 of 2 alert episodes |
No regional candidate in this study has publication approval. A zero-event or empty evaluation is insufficient evidence, not perfect accuracy. The UK archive gap prevents a current-period test; recession labels, historical definitions and sampling uncertainty limit both regions.
Regional test failures, recent windows and provenance
Euro area
Recorded archive: 2001-02-01 to 2026-09-11. Latest confirmed outcome quarter ends 2020-06-30.
Quarterly GDP growth
- Insufficient evaluated recession-onset episodes for the predeclared descriptive screen.
- Overall probability accuracy does not clear the predeclared matched-baseline screen.
- A required recent evaluation window has no eligible forecasts/outcomes.
- Too few evaluated episodes detected at the predeclared alert threshold.
| Period | Origins | Model error | Baseline error | Events |
|---|---|---|---|---|
| Since 2010 | 13 | 0.0724 | 0.0726 | 1 |
| Since 2020 | 0 | Unavailable | Unavailable | 0 |
GDP growth and momentum
- Insufficient evaluated recession-onset episodes for the predeclared descriptive screen.
- Overall probability accuracy does not clear the predeclared matched-baseline screen.
- A required recent evaluation window has no eligible forecasts/outcomes.
- A recent window performs worse than its matched training-base-rate forecast.
- Too few evaluated episodes detected at the predeclared alert threshold.
| Period | Origins | Model error | Baseline error | Events |
|---|---|---|---|---|
| Since 2010 | 13 | 0.0741 | 0.0731 | 1 |
| Since 2020 | 0 | Unavailable | Unavailable | 0 |
Input archive SHA-256: eb1a7c87b75a52bf61c96684f727cc4a1b3dc80618b16bd6cf7e9e4c01d62dbd.
United Kingdom
Recorded archive: 1976-03-31 to 2016-09-30. Latest confirmed outcome quarter ends 2016-06-30.
Quarterly GDP growth
- Overall probability accuracy does not clear the predeclared matched-baseline screen.
- A required recent evaluation window has no eligible forecasts/outcomes.
- A window exceeds the predeclared descriptive calibration error bound.
- Too few evaluated episodes detected at the predeclared alert threshold.
| Period | Origins | Model error | Baseline error | Events |
|---|---|---|---|---|
| Since 2010 | 22 | 0.022 | 0.0225 | 0 |
| Since 2020 | 0 | Unavailable | Unavailable | 0 |
GDP growth and momentum
- Overall probability accuracy does not clear the predeclared matched-baseline screen.
- A required recent evaluation window has no eligible forecasts/outcomes.
- A recent window performs worse than its matched training-base-rate forecast.
- A window exceeds the predeclared descriptive calibration error bound.
- Too few evaluated episodes detected at the predeclared alert threshold.
| Period | Origins | Model error | Baseline error | Events |
|---|---|---|---|---|
| Since 2010 | 22 | 0.0195 | 0.0189 | 0 |
| Since 2020 | 0 | Unavailable | Unavailable | 0 |
Input archive SHA-256: b7dd62a514d153396b2488503bf73f7eaba33c0284e731fa3e0d678a776dda56.
Quarterly windows overlap and the available recessions are few. Historical results have already been inspected; no untouched holdout or as-issued forecast record is claimed. Growth is computed within each available vintage to avoid splicing incompatible national-account levels.
Study regional-quarterly-fit-date-2026-10-27-v2, run with-fit-supplement, generated 27 Oct 2026, 01:12 UTC.
Protocol SHA-256: 98abf792e76bf754f716b7bfaff79604c5a0c1978cfed9a6d665fed8161638ed.
First vintage replay — superseded study design
This retained study discarded older training history already available at fitting and used a common feature start date. Its zero-onset test is superseded by the corrected study above; the original results remain available for traceability.
Historical release archive: acquired and replayed
A separate research archive now contains 65,616 observation revisions across 7 ALFRED series, including the archived recession-classification series. Source downloads, checksums and vintage dates are retained. This acquisition is additional to the legacy live model’s input records.
The as-of replay used 57,255 predictor revisions and 2,079 recession-label revisions. It produced 346 usable feature origins from 1996-12-31 to 2025-09-30. Values were selected using their recorded historical vintage availability, with an explicit conservative boundary for date-only publication information.
| As-of replay candidate | Forecast origins | Brier error | Skill against base rate |
|---|---|---|---|
| Historical-frequency baseline | 57 | 0.0171 | 0% |
| Yield-curve model | 57 | 0.1161 | -578.8% |
| Small macro model | 57 | 0.1222 | -614.3% |
Negative skill means greater squared probability error than the baseline in this sample. These performance percentages are not recession odds. The short evaluation and absent recession events prevent a general accuracy claim.
Archive coverage and remaining evidence requirements
| Series | First archived vintage | Revisions retained |
|---|---|---|
| GS10 ↗ | 1996-12-03 | 882 |
| TB3MS ↗ | 1996-12-03 | 1,115 |
| UNRATE ↗ | 1960-03-15 | 2,199 |
| PAYEMS ↗ | 1955-05-06 | 13,691 |
| INDPRO ↗ | 1927-01-26 | 39,368 |
| W875RX1 ↗ | 2010-05-28 | 6,282 |
| USREC ↗ | 2014-09-18 | 2,079 |
The first economic observation can be decades older than the first archived vintage. The replay does not pretend the archive existed before that recorded availability. Historical forecasts were reconstructed now, not issued at their stated past origins.
ALFRED vintage dates have day-level precision; the acquisition applies a documented conservative availability boundary rather than inventing exact release times. Historical USREC classifications can be provisional. Independent checks against recession announcements, current-state uncertainty, calibration and reproducibility remain necessary.
- Independent reproduction, historical label-finality review and publication approval have not been completed.
- The available as-of out-of-sample forecasts contain no recession onset; recession-detection performance is unestablished.
Study: us-asof-replay-2026-10-26-v1; generated 26 Oct 2026, 23:35 UTC. Status: insufficient evaluation events.
Protocol SHA-256: e921764d7e5ddd3df20844c923c3f4237547e6a2045f35722c6c5d6e94914871. Archive SHA-256: c083eb3e0903d54e7e4a09b26e30f95e20e5d5af1c13458e6c12fb9f39433dee.
Earlier revised-history comparison and separate IMF target
Completed US comparison study
Retrospective diagnostic only. This experiment uses revised histories and assumed release timing. It has not qualified a production probability.
365 matched forecast origins from 1993-01-31 to 2025-09-30, spanning 3 recession onsets. Each model predicts the same declared onset target on the matched dates. The historical information sets have not been proven equivalent.
| Candidate | Full-period error | Full-period skill | Since 2010 skill | Since 2020 skill |
|---|---|---|---|---|
| Historical-frequency baseline | 0.0991 | 0% | 0% | 0% |
| Yield-curve model | 0.0871 | 12.2% | -16% | -201.7% |
| Small macro model | 0.0926 | 6.6% | -29.6% | -254.2% |
| Earlier group-score model | 0.0904 | 8.8% | -19% | -163.3% |
Positive skill means lower squared probability error than this study’s historical-frequency baseline; negative skill means worse error. These percentages describe historical performance, not current recession odds. Full-period improvement does not override deterioration in recent periods.
Study limitations and reproducibility identity
- All predictors use the latest revised source histories, with an assumed two-reference-month offset. This is not real-time-vintage validation.
- Historical training labels and the condition of expansion are reconstructed from final USREC classifications; historical announcement availability is not verified.
- The protocol was fixed before these candidate results were calculated, but previous v8 evaluation results had already been inspected. No untouched independent historical holdout is claimed.
- Monthly twelve-month target windows overlap. Forecast-origin counts are not independent recession episodes; no inferential confidence interval is claimed.
- The benchmark spread is GS10 investment-basis minus TB3MS discount-basis monthly rates, explicitly distinct from the live T10Y3M spread.
- Comparisons match origin dates and target outcomes. Source sets and assumed availability differ from the archived v8 heuristic; they are not evidence of equal real-time information.
- False-alarm counts refer to alerting forecast origins, not independent alarm episodes. Lead times use month differences within each onset window.
- No hyperparameters, alert thresholds, feature choices or historical dates were retuned after inspecting this run. A better diagnostic metric does not approve a production model.
Study: us-core-diagnostic-2026-10-26-v1. Generated 26 Oct 2026, 23:35 UTC.
Protocol SHA-256: 2e54fcad51fd03826ff145f2f8a409d9516811a8bd21a0b398189eb5f5c2809d. The repository contains the fixed protocol, source manifest, calculations and full diagnostic ledger. This identifier proves the specification used, not predictive validity.
UK quarterly target foundation
A separate UK research archive now retains 2,010 quarterly GDP observation revisions across 36 provider vintage dates, plus 297 derived label revision records. These are records of a technical-GDP classification, not a count of recessions.
The source is IMF real GDP series NGDPRSAXDCGBQ, archived by ALFRED ↗. It is explicitly distinct from ONS ABMI. The longer preserved Bank of England/ESCoE archive is now integrated separately. Its ONS continuation after 2016 remains incomplete; the IMF series is not silently spliced into it.
The first available vintage is 2021-06-14, despite the underlying GDP history extending much further back. The provider’s unit metadata changes in 2024; the archive retains the source unit flags and uses within-vintage changes for the technical classification. Historical comparability and provenance still need review.
Target definition and remaining UK work
Target: UK:technical-confirmation-quarter:NGDPRSAXDCGBQ:alfred-imf-uk-quarterly-provider-native-level-v1:v1.
The separate quarterly v2 study now has a fixed protocol, chronological evaluation with recession episodes and independent automated reproduction. It uses the preserved Bank of England/ESCoE vintages; those results do not validate this IMF archive or the legacy monthly model. Its UK candidates fail the publication screens. Remaining work includes a verified post-2016 ONS continuation, review of target definitions and historical comparability, adequate recent-period and calibration evidence, and an attributable publication decision for the exact model and target. Acquiring or deriving these records does not approve a UK probability.
Archive SHA-256: 0b7230926726e6ca71136fb2223947bc2fa5d0c49ff004fced6ea687bbeb32c6. Acquired 26 Oct 2026, 23:31 UTC.
External research references
These providers answer different questions. Links are offered for comparison; their estimates are not this observatory’s approved forecast, and no provider endorsement is implied.
New York Fed yield-curve model
Estimates whether the US will be in recession in a specific month twelve months after the yield observation. This differs from a recession starting at any point during the next twelve months. A single yield-curve model can give false alarms.
Chauvet–Piger recession-state estimates
Estimates whether the US was in recession in a named historical month. These smoothed estimates can lag current conditions and change as new data arrive; they are not a forward-onset forecast.
Evidence required for publication
- Define the question. A frozen regional event, horizon and current-state condition, with a documented recession-label and revision policy.
- Establish dependable inputs. A defined core with verified geography, units, release timing, source rights and sufficient historical coverage. Additional catalogue entries remain supporting evidence unless evaluated.
- Reconstruct the information set. Audited observations and outcome labels as known at every historical forecast origin, including revisions, withdrawals and announcement delays.
- Evaluate without selecting the answer. A predeclared chronological study, target-matched simple benchmarks, untouched evaluation, recession-event counts, false alarms, lead times and calibration with defensible uncertainty.
- Approve a reproducible procedure. Independent reproduction and an attributable approval for the immutable model, region, target and refitting policy. Each forecast separately identifies its input evidence.
- Operate and retain the record. An immutable as-issued forecast ledger, live data-quality checks, independent operational monitoring, restore testing and a controlled withdrawal process.
Regional definitions and realistic limits
The US reference is NBER business-cycle dating ↗. The euro area uses CEPR–EABCN quarterly chronology ↗. The separate UK technical GDP-contraction research target now has a declared quarterly confirmation definition and provider vintage archive, as described above. It is not an equivalent monthly dating series, and its forecast model has not been validated.
A small, dependable core can be more useful for evaluation than requiring every catalogue entry. Additional indicators can strengthen context without acquiring independent numerical votes. Passing software tests, filling a missing feed or extending a recession-free test period does not establish probability calibration.
Independent review must be recorded with its actual scope. An automated evidence check is not human approval. Insufficient evidence remains an acceptable research finding, and a passing result cannot be promised by a delivery date.
Current code: concern method concern-index-1.4.1; forecast research regional-logit-1.2.0. The saved snapshot may have been produced under an earlier version.