The Yield Curve Un-Inversion Signal: A Quantitative Analysis of Recession Predictability

Abstract

This paper presents empirical evidence that the un-inversion of the U.S. Treasury yield curve (10Y-2Y spread) provides a more immediate recession signal than the widely-studied inversion itself. Across the 6 un-inversion events in our 47-year sample (1978-2024), 5 preceded NBER-dated recessions within 0-6 months (median lag: 4 months), compared to 12-24 month lead times for initial inversion. The pattern is suggestive but the sample is small: 6 independent observations provide limited statistical power, and the single false positive (2019) was followed by an exogenous pandemic recession that complicates classification. We present these results as a conditional historical pattern worthy of integration into multi-factor recession frameworks, not as a standalone predictive model.

JEL Classification: E32, E37, E44, G11, G17
Keywords: Yield curve, recession prediction, leading indicators, asset allocation, business cycles

---

1. Introduction

1.1 Research Question

Primary Hypothesis: The transition from an inverted to positively-sloped yield curve (un-inversion) provides a more immediate recession signal than the initial inversion, with recession onset occurring within 0-6 months of un-inversion.

Secondary Question: If the timing pattern holds, what should a long-horizon allocator actually do about it? We argue the answer is a change in composition rather than a change in exposure, and we set out why in Section 7.

1.2 Motivation

The 10Y-2Y Treasury yield spread inversion is widely recognized as a leading recession indicator (Estrella & Hardouvelis, 1991; Estrella & Mishkin, 1998). However, the typical 12-24 month lag between inversion and recession creates:

1. High opportunity cost: Markets often rally 15-30% during the lag period
2. Uncertain timing: Investors don't know when to act defensively
3. False persistence: Inversions can last 18+ months without immediate consequence

We propose that un-inversion solves these problems by providing a more proximate signal.

1.3 Contribution

This paper makes three contributions:

1. Empirical: Systematic analysis of un-inversion as a timing signal across all 6 events in the modern sample (1978-2024)
2. Practical: Framework for integrating un-inversion into multi-factor recession monitoring
3. Honest limitations: Transparent treatment of small-sample constraints, the 2019 ambiguity, and structural change risks

---

2. Theoretical Framework

2.1 Why Yield Curves Invert

Expectations Hypothesis: The yield curve reflects market expectations of future short-term rates (Fama & Bliss, 1987).

Inversion Mechanism: During Fed tightening cycles, short-term rates rise as the Fed raises the federal funds rate, while long-term rates fall as the bond market anticipates future rate cuts. The spread becomes negative (inversion), historically associated with recession onset 12-24 months later.

2.2 Why Un-Inversion Signals Imminent Recession

Un-inversion Mechanism: As economic weakness materializes, the Fed stops hiking or begins cutting (short rates fall), while long rates stabilize or rise slightly (reflecting term premium adjustments and shifting inflation expectations). The spread returns to positive territory — an un-inversion. Historically, this transition has coincided with recession onset within 0-6 months.

Key Insight: Un-inversion reflects the confirmation of economic weakness, not just anticipation.

2.3 Economic Mechanism: Why Un-Inversion Coincides with Recession Onset

The un-inversion is not merely a statistical pattern — it reflects the convergence of three observable economic forces that together characterize the transition from late-cycle tightening to contraction:

1. Short-rate decline driven by policy reversal. The 2-year yield falls as the Fed stops hiking or begins cutting. This is observable in the federal funds rate path and, in the post-1988 era, in fed funds futures pricing. In the recessionary un-inversions where futures data is available (2000, 2007), the market was pricing meaningful rate cuts at the time of un-inversion. For earlier episodes (1980, 1981, 1990), the mechanism operated through the same channel — the 2-year rate fell as the Fed shifted from tightening to easing — though contemporaneous futures data is unavailable.

2. Long-rate stability or rise driven by term premium. The 10-year yield does not fall as fast as the 2-year (or rises slightly) because term premium tends to increase during periods of macro uncertainty. Estimates from the Adrian, Crump & Moench (2013) model are consistent with this pattern for the post-1990 events, though term premium decomposition is inherently model-dependent and estimates vary across methodologies.

3. Credit channel transmission lag. Bank lending standards (measurable via the Fed's Senior Loan Officer Opinion Survey, available from 1990) typically tighten in the quarters preceding un-inversion, creating a contractionary credit impulse that feeds through to investment, hiring, and GDP with a lag. By the time the curve un-inverts, this impulse is already embedded in the economic pipeline.

These three forces — policy reversal, term premium repricing, and credit transmission — provide a plausible economic mechanism for why un-inversion coincides with recession onset. The mechanism is consistent with the data but should not be mistaken for a formal causal identification.

2.4 Information Content

Inversion tells you: "Recession is coming eventually" Un-inversion tells you: "The conditions that produce recessions are now materializing"

The un-inversion captures the transition from:

• Leading indicator (forward-looking expectation)
• Coincident indicator (current economic reality)

---

3. Data and Methodology

3.1 Data Sources

Primary Data:

• 10-Year Treasury Constant Maturity Rate (Federal Reserve H.15)
• 2-Year Treasury Constant Maturity Rate (Federal Reserve H.15)
• NBER Recession Dates (National Bureau of Economic Research)
• S&P 500 Total Return Index (CRSP)

Sample Period: January 1978 - December 2024 (47 years)

Frequency: Daily data, aggregated to monthly for analysis

3.2 Variable Definitions

Yield Spread: Spread = 10Y Treasury CMR minus 2Y Treasury CMR, computed daily from H.15 releases and aggregated to monthly-end values for event identification.

Inversion Event: The spread crosses from non-negative to negative territory, confirmed by the monthly-end spread remaining below zero.

Un-inversion Event (primary definition): The monthly-end spread crosses from negative to positive territory, confirmed by the spread remaining positive for at least 20 consecutive trading days (approximately one calendar month). This persistence requirement eliminates transient zero-crossings caused by daily noise. In practice, all 6 historical un-inversion events in our sample satisfied a stricter +10bps persistence threshold, so the 20-day filter does not alter the event count but would prevent false triggers from micro-crossings around zero.

Un-inversion Event (sensitivity check): We also test a +25bps threshold (spread must exceed +0.25% for at least one month) as a stricter alternative. This reduces the event count to 5 (dropping the 2019 event, where the spread briefly exceeded zero but never sustained above +0.15%) and produces a 5/5 hit rate. This stricter definition is shown only as a robustness check; it is not our primary signal. The 5/5 result should not be read as evidence of superior accuracy — it mechanically excludes the ambiguous case, which is precisely the kind of post-hoc filtering that inflates apparent performance.

Recession Proximity: Whether an NBER-dated recession start falls within 0-6, 6-12, or 12-24 months of the signal date.

3.3 Methodology

Step 1: Identify all inversion and un-inversion events (1978-2024) using the persistence-confirmed definitions above.

Step 2: Calculate time lag from each un-inversion event to the next NBER recession start date.

Step 3: Compute signal accuracy as the ratio of true positives (un-inversion followed by recession within the specified window) to total events.

Step 4: Assess what an event sample of this size can and cannot support (see Section 5).

---

4. Empirical Results

4.1 Historical Un-Inversion Events

Table 1: Un-Inversion Events and Recession Timing (1978-2024)

NBER duration sourced from NBER Business Cycle Dating Committee, "US Business Cycle Expansions and Contractions."

Observations:

• In 5 of the 6 events, an NBER-dated recession began within 0-6 months of un-inversion (median lag: 4 months). With only 6 observations, this 5/6 hit rate is suggestive but not statistically decisive — a binomial test against a 50% null yields p ≈ 0.09, which does not clear conventional significance thresholds.
• The single non-hit (2019) is ambiguous: a pandemic-driven recession followed 10 months later, raising the question of whether the signal "worked" for the wrong reason (see Section 6.2).
• The pattern is most useful as one input within a multi-factor framework, not as a standalone recession call.

4.2 Comparison: Inversion vs. Un-Inversion

Table 2: Signal Accuracy Comparison

Note on statistical power: With only 6 un-inversion events and 7 inversion events in the sample, formal hypothesis testing has limited power. A binomial test of the 5/6 un-inversion hit rate against a 50% null yields p ≈ 0.09 — suggestive but not conventionally significant. The comparison in Table 2 is best interpreted as descriptive evidence that un-inversion provides tighter timing, not as a rejection of a formal null. Larger samples (via international data or alternative spread measures) would be needed to establish statistical significance at conventional thresholds.

4.3 Case Study: 2007-2009 Financial Crisis

The 2007-2008 episode illustrates the un-inversion timing advantage:

Market impact: An investor who exited at inversion (August 2006) forfeited the +8% rally during the 14-month lag. An investor who waited for un-inversion (December 2007) captured that rally and still exited before the -57% drawdown to the March 2009 trough.

---

5. What the Historical Record Supports — and What It Does Not

The temptation at this point in a paper of this kind is to present a backtest: construct a rule, trade it across the sample, and report the risk-adjusted return. We have deliberately not done so here, and the reason is methodological rather than modest.

Six observations cannot support a performance claim. The sample contains six un-inversion events across 47 years. A strategy backtested on six signals has six degrees of freedom in its entry decisions, and essentially unlimited freedom in its exit, sizing and instrument choices. Any Sharpe ratio or drawdown figure produced from that sample is a description of six historical episodes, not an estimate of forward performance, and presenting it as the latter would invite precisely the overfitting this literature is full of. The honest statistical position is that the timing pattern is suggestive and the performance of any rule built on it is unknown.

What the record does support, and what Section 4 documents, is narrower and more useful:

• In five of six un-inversion events, an NBER-dated recession began within 0–6 months, with a median lag of four months.
• The initial inversion, by contrast, led recessions by 12–24 months — a window too wide to act on.
• The single exception (2019) was followed by a recession whose proximate cause was exogenous, which complicates classification in either direction.

That is a conditional base rate drawn from public data that any reader can verify against Federal Reserve H.15 and NBER cycle dates. It is not a strategy and we do not dress it as one.

A backtest of this framework is work we intend to publish, on an engine with point-in-time data and a survivorship-corrected universe, with the failed variants shown alongside the surviving one. Until that exists, this paper confines itself to the historical pattern and the framework for recognising it.

---

6. Economic Interpretation

6.1 Why Un-Inversion Works

Information Flow:

1. Inversion Stage (t = -18 to -6 months):

   • Market anticipates recession
   • Credit still flowing (SLOOS neutral)
   • Earnings still growing
   • Economy appears healthy (lagging data)

2. Extended Inversion (t = -6 to 0 months):

   • Credit tightening begins (SLOOS positive)
   • Business investment slowing
   • Employment weakening (but still positive)
   • Hard data starting to confirm soft data

3. Un-Inversion (t = 0):

   • Fed stops hiking (acknowledging weakness)
   • Long rates adjust to new inflation expectations
   • Spread goes positive
   • Recession imminent or underway

Key Insight: Un-inversion captures the transition from "expected" to "confirmed" recession.

6.2 The 2019 Ambiguity: Defining Signal Scope

The 2019 event requires explicit treatment because its classification depends on a definitional choice that affects the signal's reported accuracy.

Timeline:

• August 2019: Curve inverts (trade war + global slowdown fears)
• October 2019: Un-inversion
• February 2020: NBER recession begins (COVID-19)

Two valid interpretations:

1. Strict cyclical definition (our primary framing): The signal predicts recessions caused by endogenous credit/business cycle dynamics. The 2020 recession was triggered by a pandemic — an exogenous shock unrelated to the yield curve mechanism described in Section 2.2. Under this definition, 2019 is a false positive (5/6 hit rate for cyclical recessions).

2. All-NBER-recessions definition: If the signal simply predicts whether an NBER recession will occur within 12 months regardless of cause, the 2019 un-inversion was followed by a recession 4 months later. Under this definition, the hit rate is 6/6.

We adopt the strict cyclical definition because the economic mechanism (Section 2.2) describes an endogenous process — credit tightening, Fed response, hard data deterioration — that COVID did not follow. However, we acknowledge that the 2019 economy was genuinely softening (ISM <50, global PMI declining), and a milder downturn may have occurred absent the pandemic. The honest answer is that 2019 is ambiguous, and the signal's classification depends on a judgment call that cannot be resolved empirically.

Implication: Readers should evaluate the un-inversion pattern under both definitions. Neither changes the core finding that un-inversion provides tighter timing than inversion alone.

---

7. Practical Implementation

7.1 Tactical Allocation Framework

Step 1: Monitor Yield Curve Daily. Track the 10Y-2Y spread using Federal Reserve H.15 data. An un-inversion event occurs when the spread crosses from negative to positive territory.

Step 2: Hold the strategic equity weight. The instinct on this signal is to cut equity, and we think that is the wrong response. Of the ten largest daily gains since 1950, six fall within a fortnight of one of the ten largest losses — the rebound is welded to the crash. An investor who exits on a macro signal has to re-enter during the panic, and almost nobody does. Selling assets you still believe in, because the price fell, converts a drawdown into a permanent impairment. The strategic weight stays at the client's own anchor.

Step 3: Change the composition, not the exposure. A bear steepener is a repricing of duration, so the response belongs inside the sleeves rather than across them:

• Within equity — tilt toward pricing power, durable cash flow, quality balance sheets and low volatility; away from long-duration growth whose valuation depends on a low discount rate. Rising long rates compress the latter most severely, by construction.
• Within fixed income — shorten duration. A bear steepener punishes the long end mechanically. Long bonds here are not a hedge against the equity position; they are a second position in the same risk.
• The haven sleeve — gold and inflation-linked bonds earn their place when long rates rise against restrictive policy: one works if the impulse proves inflationary, the other if it does not. Both are funded from the bond and cash bucket, never by cutting equity.

Step 4: Rebalance into weakness. Holding the anchor through a drawdown means buying equity as it falls. This is the uncomfortable part and it is also the part that works. Rebalancing toward a falling asset is the part of the discipline that gets abandoned first, which is why it has to be written down in advance rather than decided in the moment.

A note on governance. A framework that holds the strategic anchor is materially easier to operate inside an investment policy statement than one that does not: there is no tracking-error breach requiring an exception, no committee debate about when to re-enter, and no discretionary de-risking decision to defend in a review three years later. The discipline sits in the rotation rather than in the timing.

7.2 Integration with Other Indicators

Multi-Factor Context:

The un-inversion signal is most useful when combined with corroborating indicators. A multi-factor framework might include:

When multiple indicators align with the un-inversion signal, historical precedent suggests elevated recession risk. When they diverge (as in 2019, when the Fed cut preemptively and credit conditions remained loose), the signal is less reliable.

Note: We deliberately avoid publishing a specific composite probability or recommended allocation. The weights and thresholds for combining these indicators are implementation-dependent and should reflect each investor's risk tolerance, time horizon, and asset mix.

---

8. Limitations and Future Research

8.1 Sample Size

Limitation: Only 6 un-inversion events in 47 years

Impact: Statistical power is limited

Mitigation:

• Use international data (Germany, UK, Japan yield curves)
• Study state/local yield curves (municipal bonds)
• Extend analysis to credit spreads (Baa-Aaa)

8.2 Structural Changes

Concern: Fed's QE programs distorted yield curve (2008-2022)

Analysis:

• Un-inversion still worked in 2008 (during QE1)
• 2019 false positive coincided with "NOT QE" (Fed balance sheet expansion)

Conclusion: QE may reduce inversion reliability, but un-inversion appears robust

8.3 Future Research

1. Cross-Country Analysis: Do un-inversions predict recessions in Europe/Asia?
2. Corporate Credit: Do Baa-Aaa spread un-inversions provide similar signals?
3. Machine Learning: Can neural networks improve timing by 1-2 months?
4. Behavioral Finance: Why do investors ignore un-inversion signals?

---

9. Conclusion

This paper documents a historical pattern: in the modern sample (1978-2024), yield curve un-inversion has provided tighter recession timing than the widely-followed inversion signal. Our key findings:

1. Timing: In 5 of 6 events, recessions began within 0-6 months of un-inversion (median 4 months)
2. Sample constraint: 6 observations provide suggestive but not statistically decisive evidence — the pattern should inform, not dictate, allocation decisions
3. No performance claim: we deliberately present no backtest — six observations cannot support one (Section 5)

Practical Implication: Investors should monitor yield curve transitions (inversion → un-inversion), not just absolute levels. The un-inversion is most valuable as one input within a multi-factor recession framework.

Current Application (December 2025): The yield curve un-inverted in September 2024. Conditional on the historical pattern, this places the current environment in the window where recession risk has historically been elevated (0-14 months post-un-inversion). This does not constitute a recession forecast — it means the historical base rate of recession following this signal is higher than unconditional probabilities would suggest, and risk management should reflect that asymmetry.

---

References

Adrian, T., Crump, R. K., & Moench, E. (2013). Pricing the term structure with linear regressions. Journal of Financial Economics, 110(1), 110-138.

Engstrom, E. C., & Sharpe, S. A. (2019). The near-term forward yield spread as a leading indicator: A less distorted mirror. Financial Analysts Journal, 75(4), 37-49.

Estrella, A., & Hardouvelis, G. A. (1991). The term structure as a predictor of real economic activity. Journal of Finance, 46(2), 555-576.

Estrella, A., & Mishkin, F. S. (1998). Predicting U.S. recessions: Financial variables as leading indicators. Review of Economics and Statistics, 80(1), 45-61.

Fama, E. F., & Bliss, R. R. (1987). The information in long-maturity forward rates. American Economic Review, 77(4), 680-692.

Federal Reserve Board. (2024). Selected Interest Rates (H.15). Retrieved from https://www.federalreserve.gov/releases/h15/

National Bureau of Economic Research. (2024). US Business Cycle Expansions and Contractions. Retrieved from https://www.nber.org/cycles.html

---