← All Projects
Monetary Policy Inflation Rates

Yield Curve Inversions as Recession Predictors: A Post-QE Reassessment

Published March 2025
Reading time ~18 min
Data through Q4 2024
Charts Interactive · 5
Abstract

The 2s10s Treasury spread inverted in March 2022 and remained negative for the longest sustained period on record — yet a textbook recession has not followed. This paper re-examines the 2s10s as a leading indicator in the context of a post-QE term structure, arguing that balance sheet expansion has durably compressed the term premium and mechanically biased the curve toward inversion. We propose an adjusted spread that strips out the estimated term premium contribution, and find its predictive record remains broadly intact once this correction is applied.

1. Introduction

The slope of the yield curve has long occupied a central place in macroeconomic forecasting. An inverted curve — where short rates exceed long rates — has preceded every US recession since the 1960s, typically with a lag of 12 to 18 months. The mechanism is intuitive: inversion signals that markets expect the central bank to cut rates in the future, usually in response to economic weakness.

The 2022–24 inversion complicated this story. The Federal Reserve embarked on its fastest tightening cycle since the early 1980s, pushing the 2-year yield sharply above the 10-year. Yet GDP growth remained positive, the unemployment rate stayed near historical lows, and corporate spreads never reached distressed levels. Was the recession signal wrong, or was the indicator itself broken?

This paper argues the latter — but not irreparably so. Unconventional monetary policy has left a structural residue in long-end yields: a persistently suppressed term premium that biases the 10-year rate downward relative to where it would price absent QE. Stripping out this distortion restores much of the curve's informational content.

2. Data & Methodology

2.1 The raw 2s10s and its historical record

We use daily Federal Reserve H.15 data on constant-maturity Treasury yields from January 1976 through December 2024. NBER recession dates are used to define the dependent variable. An inversion is defined as any period where the 2s10s spread is negative for at least 10 consecutive business days.

Fig. 1 — US 2s10s Treasury Spread (1990–2024) & Recession Periods Source: Federal Reserve H.15 / NBER

2.2 Term premium decomposition

We adopt the Adrian, Crump & Moench (ACM) term premium estimates published by the New York Fed. The ACM model decomposes the 10-year yield into the risk-neutral rate — the expected path of short rates — and the term premium, which compensates investors for bearing duration risk.

Our adjusted spread is defined as:

Definition

Adjusted Spread = (10Y Yield − ACM Term Premium) − 2Y Yield

This isolates the expectational component of the long end, removing the mechanical compression introduced by QE-driven term premium suppression.

2.3 Recession probability model

We estimate probit models using (i) the raw 2s10s, (ii) the adjusted spread, and (iii) a multivariate specification adding the unemployment rate gap and ISM Manufacturing PMI. Models are estimated over 1976–2019 and evaluated out-of-sample over 2020–2024.

3. Key Findings

Table 1 below summarises the lead-time and false-signal record of the raw versus adjusted spread across all inversion episodes since 1978.

Inversion episode Duration (months) Recession followed? Lead (months) Adj. spread signal
Aug 1978 – Oct 1980 26 Yes (×2) 6 / 18 Confirmed
Sep 1980 – Oct 1981 13 Yes 9 Confirmed
Dec 1988 – Mar 1990 15 Yes 14 Confirmed
Feb 2000 – Jan 2001 11 Yes 13 Confirmed
Aug 2006 – Jun 2007 10 Yes 16 Confirmed
Mar 2022 – Jan 2024 22 No (so far) Not confirmed
Fig. 2 — Raw vs. Adjusted 2s10s Spread (2018–2024) Source: Federal Reserve H.15 / NY Fed ACM Model

The most striking result is that the adjusted spread never inverted below −50 bps during 2022–24, compared to a trough of −108 bps for the raw 2s10s. In all prior episodes that were followed by recessions, the adjusted spread breached −75 bps. By this measure, the 2022 cycle transmitted a warning — elevated risk — without triggering the historical alarm threshold.

Key Finding

The 2022–24 inversion reflects a term premium deficit of approximately 120–150 bps relative to pre-QE norms. Correcting for this, the adjusted spread peaked at −42 bps — consistent with elevated but not recessionary risk conditions, in line with the observed soft-landing outcome.

4. Probit Model Results

The out-of-sample performance of the three specifications over 2020–2024 is summarised in Fig. 3. The raw 2s10s probit assigns a peak 12-month recession probability of 74% in Q3 2023. The adjusted spread model peaks at 38%. Neither triggered. The multivariate model — which includes the unemployment gap — assigns a more conservative 28% peak, the closest to realised outcomes.

Fig. 3 — 12-Month Recession Probability (Probit Models, 2020–2024) Source: Author calculations

These results do not invalidate the yield curve as a signal — they suggest the appropriate conditioning set has expanded in the QE era. Labour market resilience, excess household savings, and a strong starting point for corporate balance sheets all operated as offsets that the raw curve alone cannot capture.

5. Conclusion

The yield curve remains a valuable leading indicator, but its signal must be interpreted through the lens of a post-QE term structure. A durably compressed term premium means that observed inversion understates the true policy tightness implied by market expectations. Once this is corrected, the 2022–24 episode is consistent with elevated cycle risk — but not the outright recessionary signal the raw spread implied.

Going forward, the ACM-adjusted spread should be the primary curve indicator for recession forecasting. Analysts relying on the raw 2s10s face a structurally biased signal for as long as central bank balance sheets remain elevated relative to pre-2008 norms.

Data: Federal Reserve H.15 (constant-maturity yields); NY Fed ACM term premium model; NBER Business Cycle Dating Committee. All calculations are the author's own. Model files available for download above.

Related Projects All projects →