Two channels, separated by construction
Credit market variables enter only the credit equation and oil price variables only the oil equation, and the eight postwar episodes are assigned to channels on narrative grounds. The separation gives each coefficient an economic reading and lets channel identification be tested against labels that do not embed the predictors; the model matches the narrative classification in six of eight episodes.
Persistence against lead time
The lagged recession indicator is the largest term in either equation, and the probability formed inside a recession exceeds 50 percent in 33 of 35 recession quarters. One quarter before onset the signal is below 11 percent in seven of eight episodes, and looking forward from an expansion quarter the twelve-month AUC is 0.802, against 0.935 in sample.
The line of sight
Two rails run down each section’s left edge: the left in the color of the swatch of “Two channels, separated by construction,” the right in that of “Persistence against lead time.” A filled mark ( ) names a section that carries the theme; a hollow mark ( ) a theme that passes through it. Each section names its themes in words as well.
1
Introduction
- Establishes
- The eight postwar recessions that the National Bureau of Economic Research (NBER) dates divide on narrative grounds into credit-channel episodes (1969–70, 1981–82, 1990–91, 2001, 2008–09), preceded by a deterioration of credit market conditions as the EBP measures it, and oil-channel episodes (1973–75, 1980), preceded by oil price shocks; 2020 is assigned to the oil category as the supply-side outcome class. Under the exclusion restrictions each coefficient carries a direct economic reading, and the model cannot conflate the two channels.
- Leads to
- Which literatures does the model rest on, and what has each established about the predictor it contributes?
2
Literature Review
- Establishes
- The work of Estrella and Hardouvelis and of Estrella and Mishkin established the term spread as a predictor of U.S. recessions, and Johansson and Meldrum show that lower term premiums let the yield curve flatten without a matching rise in recession risk. The term spread’s predictive content has weakened: an inversion had reversed by the onset quarter of 2008–09, and the persistent inversion of 2022–23 was followed by no NBER-dated recession. Dueker introduced the lagged recession indicator, and Kauppi and Saikkonen developed a framework of four specifications, whose dynamic form the paper adopts. Gilchrist and Zakrajšek decompose a credit spread index constructed from micro-level corporate bond data into expected default and the EBP, the predictive content for activity lying in the EBP. Hamilton documents that oil price increases preceded most postwar recessions and that the relation is nonlinear in large positive shocks.
- Leads to
- Which series does the model use, how is each built from its primary source, and on what grounds is each recession assigned to a channel?
3
Data
- Establishes
- Every series comes from its primary release: the term spread as the 10-year Treasury yield minus the 3-month Treasury bill rate, the EBP from the Gilchrist–Zakrajšek file, the Baa–Aaa spread from Moody’s, the crude oil price from the producer price index for crude petroleum, and real gross domestic product (GDP) growth, each monthly series aggregated to the quarter by the mean of its three months. The EBP begins in 1973:Q1; earlier quarters carry a synthetic series, the Baa–Aaa spread standardized to the EBP distribution over the 214-quarter overlap, so the 1969–70 result rests on a proxy. The channel labels are assigned on the proximate cause documented in the business cycle literature, not from the predictor values; 1990–91 and 2020 carry stated caveats, 2020 fitting neither channel and assigned to the oil category as the supply-side outcome class.
- Values
- 245Quarters in the sample, 1965:Q2 through 2026:Q235Recession quarters in the sample214Quarters of overlap behind the synthetic EBP
- Exhibits
- Tables 1 and 2
- Leads to
- How does a three-category model keep the two channels apart, and how does it carry recession persistence?
4
Methodology
- Establishes
- In the multinomial logit, expansion is the base category, and two index functions carry the risk of the two recession types: the credit index only credit market and term spread variables, the oil index only oil price variables, the exclusion restrictions enforced by construction. The Kauppi–Saikkonen extension adds the lagged recession indicator to both index functions, so the forecast for the next quarter conditions on the current quarter’s regime. The EBP, the residual of a credit spread index decomposed against a Merton distance-to-default model, reads the risk-bearing capacity of the financial sector. It stood at +3.27 pp in 2008:Q4 and at +1.59 pp in 2000:Q4, one quarter before the 2001 recession, and it was negative before the 1973–75, 1980, and 2020 episodes, a pattern consistent with the exclusion restrictions.
- Values
- +3.27 ppEBP in 2008:Q4, inside the 2008–09 recession+1.59 ppEBP in 2000:Q4, one quarter before the 2001 onset
- Exhibits
- Equations (1) and (2)
- Leads to
- What does each of the two model changes add, measured on its own?
5
Model Progression
- Establishes
- Holding the credit variable at the Baa–Aaa spread, the dynamic extension raises the AUC from 0.890 to 0.926 and F1 from 0.582 to 0.783, and lowers the Brier score from 0.090 to 0.053. Replacing the spread with the EBP moves every metric and both information criteria the same way, the AUC to 0.935. The gain concentrates in the 2001 episode: the signal one quarter before onset rises from 4.7 percent to 45.0 percent, the EBP having stood at +1.59 pp while the Baa–Aaa spread stood at 0.81 pp, the expected-default component masking the intermediation signal.
- Values
- 0.890AUC, static model with the Baa–Aaa spread0.926AUC, dynamic model with the Baa–Aaa spread0.935AUC, dynamic model with the EBP4.72001 signal one quarter before onset, dynamic model with the Baa–Aaa spread, percent45.0The same signal, dynamic model with the EBP, percent
- Exhibits
- Table 3
- Leads to
- What are the coefficients of the preferred model, and what does it signal before an onset and inside a recession?
6
Results
- Establishes
- In the credit equation the EBP enters the current positive portion with a coefficient of 2.309 and its four-quarter average with −0.182, so the contemporaneous reading carries the credit signal, and the lagged recession indicator enters with 3.840, the largest term in the equation. The two term spread terms enter with opposite signs, and a sustained inversion of 1 pp lowers the credit index by 0.49: conditional on the EBP and on persistence, the term spread does not reproduce the inversion mechanism of the yield curve literature. The model identifies the channel consistent with the narrative classification in six of eight episodes, the exceptions being 1981–82, where the two channel probabilities differ by 0.1 percentage point, and 2020, where neither indicator carried a signal. The probability formed in a recession quarter for the next exceeds 50 percent in 33 of 35 quarters; after the 2008–09 episode the forecast formed in 2009:Q3, the first expansion quarter, when the lagged indicator returns to zero, declines to 1.0 percent.
- Values
- 3.840Coefficient on the lagged recession indicator, credit equation2.309Coefficient on the positive portion of the EBP33 of 35Recession quarters signaled above 50 percent49.91982:Q3, a quarter below the threshold, percent26.02020:Q2, the other quarter below the threshold, percent
- Exhibits
- Tables 4 and 5; Figure 1
- Leads to
- Does the discrimination survive outside the estimation sample, and against which benchmarks?
7
Out-of-Sample Evaluation
- Establishes
- An expanding-window exercise re-estimates the model each quarter from 1990:Q1, giving 146 one-quarter-ahead predictions over four recessions. The AUC declines from 0.935 in sample to 0.925, both computed over every prediction, the recession quarters among them. A probit on the term spread alone reaches 0.547 over the same window, close to the 0.5 of an uninformative ranking, and its Brier score of 0.104 exceeds that of a rule that calls a recession on negative GDP growth: over this window the term spread on its own does not order recession risk. Three features imply that the reported performance bounds real-time performance from above: predictors enter at their most recent vintage, the lagged recession indicator is taken from the final NBER chronology, and the F1 thresholds are selected ex post.
- Values
- 0.925AUC out of sample, 146 predictions from 1990:Q10.547AUC of the term-spread probit over the same window0.104Brier score of the term-spread probit
- Exhibits
- Table 6
- Leads to
- What does the model forecast from the most recent quarter, one quarter and twelve months ahead, and why is its response to a 53.3 percent rise in crude oil prices small?
8
Current Forecast
- Establishes
- As of 2026:Q2, the model forecasts for 2026:Q3 expansion at 94.2 percent, a credit-channel recession at 3.5 percent, and an oil-channel recession at 2.3 percent. The producer price index for crude petroleum stands 53.3 percent above its year-earlier level, comparable to the readings before the 1980 and 1981–82 episodes, while the EBP is negative at −0.29 pp and the term spread positive at 0.80 pp. The oil-channel response is small for two reasons the specification makes explicit: the oil equation weights the four-quarter average, which stands at 7.1 percent because the three preceding quarters were flat or negative, and the lagged recession indicator, the largest term in either equation, is zero. On the twelve-month definition that published recession probabilities use, the probability of a recession quarter in the four quarters after 2026:Q2 is 21.2 percent, against 17.9 percent a quarter earlier; the credit channel’s share fell from 15.2 to 12.7 percent and the oil channel’s rose from 2.7 to 8.5 percent. The panel of the Wall Street Journal survey of October 2026 put the probability at 24 percent. Over the 206 expansion origins whose window is fully observed, the twelve-month figure reaches an AUC of 0.802: looking forward from an expansion quarter, the model orders recession risk at 0.80 rather than at the in-sample 0.94.
- Values
- 94.2Expansion probability for 2026:Q3, percent21.2Recession quarter in the four quarters after 2026:Q2, percent8.5The oil channel’s share of it, percent53.3Crude petroleum price index above its year-earlier level, percent
- Exhibits
- Tables 7 and 8; Figures 2 and 3; Equation (3)
- Leads to
- Which findings does the paper carry forward, and what do they not support?
9
Conclusion
- Establishes
- Four findings. The dynamic extension raises the AUC from 0.890 to 0.926 with the Baa–Aaa spread held in place and supports within-recession signaling at 33 of 35 quarters above 50 percent. Replacing that spread with the EBP raises the AUC to 0.935 and the 2001 pre-onset signal from 4.7 percent to 45.0 percent, by isolating the financial intermediation channel from expected default risk. Channel identification matches the narrative classification in six of eight episodes. The term spread does not reproduce the inversion mechanism once the EBP and persistence are conditioned on, and a term-spread probit reaches an out-of-sample AUC of 0.547: in this specification the credit-channel signal is carried by the EBP. On the twelve-month definition the reading after 2026:Q2 is 21 percent, against 18 percent a quarter earlier. The model’s limitations follow from its structure: the signal one quarter before onset is below 11 percent for seven of the eight episodes, and the channel classification is a narrative assignment. The delay of the NBER’s announcements and the final-vintage data imply that reported performance bounds real-time performance from above, as Section 7 states.