Skip to content

Backtest

The same methodology run month by month from 1970 using only data dated before each month, evaluated against NBER business-cycle dates. Parameters were chosen in advance; the sensitivity table shows results across thresholds instead of selecting the best one.

Parameters

Evaluated 1981-09-30 → 2026-09-30. A detection = signal confirmed between 18 months before an NBER peak and the trough. A false positive = a confirmed signal with no recession in that window. Lead = months from first confirmed signal to the peak (negative = after the peak).

4
Recessions evaluable
3
Detected
1
False negatives (missed)
4
False positives
1.3
Average lead (months)
1
Median lead (months)
7%
Expansion months ≥ threshold
0
Unresolved recent signals

Score history vs NBER recessions (shaded)

Per-recession results

RecessionEvaluableDetectedFirst signalLead (m)Max 12m beforeAt peakCoverage
1969–70No (history)——————
1973–75No (history)—————34%
1980No (history)—————44%
1981–82No (history)—————44%
1990–91YesYes1990-08-1606264%
2001YesYes2000-114636573%
2007–09YesYes2007-111636478%
2020YesMissed——454383%

Signal episodes

ConfirmedEndedPeak scoreClassification
1981-101982-1182Preceded/coincided with 1981–82
1990-081991-0678Preceded/coincided with 1990–91
2000-112000-1262Preceded/coincided with 2001
2001-032002-0467False positive
2007-112009-1280Preceded/coincided with 2007–09
2020-052020-0765False positive
2023-042023-0562False positive
2023-102023-1162False positive

Score around NBER peaks (event study)

Score behaviour before / during / after

PhaseAverage scoreMonths
12 months before peak5048
During recession7251
12 months after trough5260
Other expansion months37382

Threshold sensitivity (not an optimisation)

ThresholdDetectedMissedFalse pos.Avg lead
4540410.0
503128.0
553145.0
603141.3
65310-4.0
70220-3.5
75220-6.0

Caveats — read before drawing conclusions

  • Uses today's revised data from FRED, not the data available in real time (no vintage/ALFRED data). Revisions can make historical signals look cleaner than they were.
  • Publication lags are approximated (daily 1d, weekly 7d, monthly 40d, quarterly 120d).
  • Percentiles are computed point-in-time (expanding window, minimum 5 years of history), so no future data enters any historical score.
  • Indicator coverage varies over time: several series (ICE credit spreads, breakevens, JOLTS, VIX, SLOOS) start decades after others. Months below the minimum coverage are excluded.
  • Only ~7-8 recessions are available. Weights and thresholds were set judgementally beforehand and are NOT optimised to these episodes; the sensitivity table shows how results vary with the threshold rather than picking the best one.
  • A signal is defined as the score staying at or above the threshold for the required consecutive months, dated when that is first confirmed. A detection counts if the signal is confirmed between the horizon before the NBER peak and the trough.