The abstract of a new paper coauthored with Laurent Ferrara, “Missing Recessions: Revisiting the Term Spread’s Predictive Power”:
The failure of a recession to materialize in 2024 has forced a reappraisal of the term spread as the most reliable recession predictor for the US. In this paper, we re-evaluate the predictive power of the term spread, and explain the false positive signal in terms of sample specificity, simultaneity, and omitted variables, focusing on the United States, but including other advanced economies for which the term spread has proven to be a useful predictor in the past. We confirm that prior to the 2020 recession, the yield curve was an accurate predictor of recessions in the US, Canada, and Germany, but proved less accurate thereafter. For the US, accounting for term premia does not improve prediction appreciably, while taking into account expectations of future unemployment does. More definitively, in line with Borio, et al. (2020), we find inclusion of the debt service ratio reconciles the failure of a recession to appear in 2024 after the strong yield curve inversion observed at the end of 2022, for both the US and Germany.
Focusing on the United States, it’s clear that most term spread measures (10yr-3mo, 10yr-2yr, etc.) indicated a high likelihood of recession in 2023-24, given the large and persistent inversion in 2022. Relying on pre-2020 data, recession probabilities looked like this.
Figure 1: Top panel – US 10yr-3mo spread, %; bottom panel – probability of recession, using regression estimated up to 2019M12. NBER defined peak-to-trough recession dates shaded gray.
There are many possible explanations for the term spread mis-fire of 2024: the data generating process has changed since 2020, the behavior of the term premium (which is part of the term spread) has changed, omission of Taylor Rule variables like expected unemployment results or omission of financial variables like the financial cycle have biased the estimates. It’s also possible that “by chance” the recession just didn’t occur; after all, one is never 100% confident.
In this paper, we examine these possibilities, by estimating term spread specifications over a full sample, or augmenting with key variables.
Source: Table 2 from Chinn and Ferrara (2026).
Note that the proportion of variance explained using the term spread is higher in the pre-2020 sample, and lower in the full (0.33 vs. 0.26). Adding in a separate estimated term premium (the ACM estimate from the NY Fed) does not appreciably change the full sample result. However, adding in the expected change in unemployment rate (because the Taylor rule implies the Fed funds expected in the future will be lower if markets anticipate a slowdown/recession) does result in a marginally significant coefficient for this variable (although the pseudo R2 is essentially unchanged).
Only inclusion of a proxy for the financial cycle — the debt service ratio (DSR) — raises the proportion of variance explained appreciably, to 0.57. It also turns out that the best “fit” — in the sense of no false positives, while catching the recessions of 1990-91, 2008-09, is the equation incorporating the DSR (teal line in figure below, marked &DSR).
Source: Figure 5 from Chinn and Ferrara (2026). NBER defined peak-to-trough recession dates shaded gray.
The finding that the DSR’s inclusion helps in accurately predicting recessions holds true even if one restricts the estimation period to pre-2020. Then the 1990-91, 2001, and 2008-09 recessions are predicted, but not the 2020 pandemic one (nor a 2024 recession).
It’s not that accounting for the Taylor rule is completely useless (column 4 in Table 2). Including the expected change in the unemployment rate (conjoined with using the Fed funds rate instead of 3 month rate) reduces the estimated probability of recession for 2024 below the 40% threshold (blue line, marked &UE), while catching previous recessions, save the 2008-09 one. So in that sense, not explicitly taking into account the Taylor rule could be a part of the explanation for a recession call for 2024.
Read the entire paper here, which also examines the experiences of Canada and Germany. This paper is a substantially revised version of paper described here):
Thank you for this. I do have a couple of questions/comments.
The hypothesis in this paper appears to be something of a mirror image of my hypothesis that the continuing effects of fiscal stimulus (or austerity) can be enough to overcome the “message” of the simple yield curve.
In that regard, I wonder if you are able to take this analysis back to the 1960s to account for the ‘other’ false positive yield curve inversion of 1966? Or is the data simply not available?
Secondly, I did read your paper and also skimmed Borio et al’s 2020 paper similarly suggesting that financial conditions are a more potent metric of expansion vs. recession than the yield curve, as well as your earlier note, with included graph of the historical debt service ratio. It appears the data you have used is similar to although not quite the same as that in the below FRED graph:
https://fred.stlouisfed.org/graph/fredgraph.png?g=1ZJk4&height=490
In addition to being curious about the difference, could you explain in layperson’s terms what change in the debt service ratio is used to generate your outcomes? Most notably, there was a big increase in the DSR in 2022-23 compared with the stimulus and near “0” interest rate years of 2020-21, but the 2022-23 levels were nevertheless much lower than any levels before the pandemic, which leads me to think you are utilizing a comparison over a similar longer term period.
Thanks for your anticipated response.
NDd,
I think there may be a bridge between your notion that financial conditions – which are closely related to debt expansion – are positively correlated with growth and that debt service is negatively correlated. I doubt I need to spell it out, but today’s stimulative expansion comes back to haunt as tomorrow’s debt service. It’s a pattern that has haunted failed regimes and imposed lesser costs even on successful regimes.
Once the debt/income ratio becomes large, current-period debt expansion can drive up current-period debt service with interest-rate increases. That’s relevant today, but not always.
Hey, isn’t there another guy who has worked with Ferrera on the debt-service ratio as a variable in recession identification?:
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4709138
I would expect that you cannot have a recession without an issue with employment. so how does the term spread correlate with employment? in the 2024 case, employment never really became a big issue. has that ever been the case in the past, when the term spread predicted the recession? I guess this is the problem when you want a simple one variable model. that single variable may be a dominant feature, but not the whole story. so do you want complex and accurate models, or simple and pretty good models, in real time?