Inverted Yield Curve, No Recession: What a Real-Time Model Showed, 1976-2026

Latest observation

Summary

The real-time record in six numbers

What the probability is and is not

The line in the chart at the top is the probability of being in recession 12 months later. Formally, it is the model’s estimate that USREC equals 1 in month t+12, given the spread in month t. It does not give the probability that a recession starts at some point in the next 12 months. On the in-sample replication described below, that second target fits less well, with a pseudo-R² of 0.072 against 0.137 for the t+12 target (408 monthly origins).

The reading has one input. The spread, explained in what the 10-year/3-month Treasury spread measures, enters a probit with two coefficients that are re-estimated each month. For a given month’s coefficients, the probability depends on the spread alone. Across the 7 episodes in the evaluation, the peak probabilities rank exactly as the spreads at those peaks (Spearman −1.000). The model adds a calibration to the spread and leaves the order of the episodes unchanged.

Calibration is where the model earns or loses its keep. In the 41 months where it would have shown 0.5 or more, the average reading was 0.707, and the share of those months followed by a recession month 12 months later was 0.268. In the 0.3 to 0.5 band the two were near each other, 0.395 against 0.419 over 43 months.

Base model (12-month information lag), evaluation filter. Each band includes its lower bound and excludes its upper bound.

The steep end of the curve rests on few months. Since 1953 the spread has been below −1.50 pt in 8 months: 7 fell in 1979, 1980 and 1981, the eighth in 2023. A reading of 0.766 at −1.57 pt in May 2023 reflects what the model learned from those three years.

How to read the chart: the line is the reading each month would have produced, the grey bands are USREC recession months, and the horizontal line marks 30%, the threshold crossed before each of the 6 evaluated recessions. The marker sits on May 2023, the peak of the 2022-24 cycle.

The edge the model built from 1976 to 2021

The case for the curve rests on a long record. Eco3min’s Q&A on why the yield curve inverts and its comparison of how accurate recession indicators have been both cover the curve’s record before recent US recessions. This study does not revisit that history. It asks what a model fitted only on data available at each date would have shown.

One replication comes first. The analysis of the New York Fed recession probability model reports an in-sample pseudo-R² of 15.6% on 408 months. Fitted on this study’s monthly file over 1990-2024, the same probit gives 0.137 with the t+12 target (31 recession months among 408) and 0.143 on the bond-equivalent spread. The 2-month inversion rule finds 3 true positives and 1 false positive (2022-11) over the same years, matching the four sustained episodes that page counts. Its six true positives come from daily T10Y3M data, where re-inversions count separately. The remaining gap is consistent with a different 3-month series and a different way of counting the target, and the T10Y3M probit walkthrough sets out the in-sample version. From here on, the study uses out-of-sample readings only.

The out-of-sample test begins in January 1976. At each date t, the probit uses only pairs whose outcome month falls at least 12 months before t, a stand-in for the delay in dating recessions. Split at the publication of this model in 1989-1991, the record holds on both sides. From 1976 to 1990, before journal publication, the Brier skill score against the base rate was +0.251 (90% interval −0.046 to +0.555, 30 recession target months). From 1991 to 2021, with the specification already in print, it was +0.212 (+0.079 to +0.447, 28 target months). The Diebold-Mariano test rejects equal accuracy in that second segment (t 2.07, p 0.019).

Base model, 12-month information lag. The full evaluation covers 6 recessions. No skill score is published for the 2022-24 window alone, which has no recession target: the cumulative edge chart below shows that period instead.

Threshold readings tell the same story. Every 30% crossing before 2022 was followed by a recession start within 24 months: 8 true positives, covering all 6 evaluated recessions, with a median lead of 13.0 months (9 to 23). The 2020 recession is an exogenous shock, the pandemic, counted by protocol. Dropping origins from 2018 to 2020 raises the full-sample score from +0.054 to +0.073.

What 2022-24 did to that edge

This study started from a different claim, that the May 2023 reading stood above the reading before every recession on record. The data do not support it. At 0.766 and −1.57 pt, May 2023 is third of 7 episode peaks, below the readings ahead of the 1981 recession (0.992 in December 1980, at −2.65 pt) and the 1980 recession (0.920 in December 1979, at −1.65 pt). It stands above the peaks ahead of the 2001 (0.492 at −0.53 pt), 2008 (0.467 at −0.38 pt), 2020 (0.446 at −0.32 pt) and 1990 (0.304 at +0.13 pt) recessions. What holds in all five variants is the erosion of the cumulative edge.

The sequence ran as follows. The base model crossed 30% in November 2022 (0.406 at −0.26 pt) and 50% in December 2022 (0.513 at −0.63 pt), then peaked in May 2023. It crossed 50% again in November 2023 (0.555 at −0.77 pt) and still read 0.344 in November 2024, the last negative month (−0.06 pt). USREC recorded no recession month in that window or after it, through October 2026.

The cumulative edge over the base rate, 1976–2025

Brier score of the known recession frequency minus Brier score of the model, summed over monthly origins. Above zero, the model scored better.

Source: FRED (GS10, TB3MS, USREC). Eco3min Research calculation on the published CSV.

Read the chart from left to right. The base model’s cumulative gap was 6.48 at the end of 1990, 9.95 at the end of 2001, 11.44 at the end of 2009 and 12.48 at the end of 2021, with its high at 13.65 in April 2019. By October 2025 it was 2.95, so 76.4% of the end-2021 edge was gone. The scoring is asymmetric. A confident reading followed by no recession costs more squared error than a calm month earns, so a few months above 0.5 undid years of accumulated edge.

The mechanism is the depth, and the depth came from the short end. Between November 2020 and May 2023 the 3-month bill rose 5.05 points and the 10-year yield 2.70 points, so the spread fell 2.35 points to its trough. In the three inversions before it, the long yield moved little or fell: −0.10 point before the 2000 trough, +0.28 before 2007 and −0.56 before 2019. A probit sees only the difference, so it read the 2023 trough like the troughs of 1979 to 1981 it had been fitted on. The transmission from an inversion to a recession, through bank lending margins and credit supply, is set out in the credit-channel analysis, and the model observes none of it.

The case for a late signal

The strongest reply to this record is that the recession is late, and that one episode settles little. Both points have support in the data. The sample holds 6 evaluated recessions and 58 recession target months, and the full-sample interval runs from −0.326 to +0.308, so a score of zero and a score of +0.3 are both consistent with what has been observed. The argument that a slow signal keeps its value, made in why a slow regime signal still works, fits that width.

Two tests separate what the data can and cannot say. The Diebold-Mariano test asks whether the model’s squared errors are smaller than the base rate’s on average. Over the full evaluation it does not reject equality (t 0.31, p 0.378). The Clark-West test asks a narrower question, suited to nested models: does adding the spread to a constant-probability model carry information, once the noise of estimating the extra coefficient is allowed for? It does, with t 2.90 and p 0.002. The spread carries information beyond the base rate, while the gain in accuracy over the full window is not statistically distinguishable from zero.

The robustness checks, in full. Shortening the information lag to 6 months or lengthening it to 18 moves the score by −0.003 and −0.002 (paired 90% intervals −0.013 to +0.009 and −0.007 to +0.002). Measuring the bill on a bond-equivalent basis, the New York Fed convention, raises it by +0.038 (+0.000 to +0.095). A 20-year rolling window raises it by +0.046 (−0.056 to +0.183) and erases the least of the edge, 50.2%. A control at a 24-month lag, outside the published variants, gives +0.049 and a 2022-24 peak of 0.769 (May 2023, −1.57 pt). Across the five published variants, the share of the end-2021 edge erased by October 2025 runs from 50.2% (20-year rolling window) to 77.1% (6-month lag).

All five variants are scored on the same 598 origins and 6 recessions. The rolling window estimates each month on the previous 240 pairs only, so by 2023 its sample no longer includes the inversions of 1979 to 1981, and its 2023 peak is 0.604.

Three objections from the other side deserve their full weight. The first is the term premium. A 10-year yield held down by central-bank bond holdings can invert the curve without the usual message about expected rate cuts, and a two-series model cannot tell the two cases apart. The second is the short leg: the 2022-24 inversion rested on a 5.05-point rise in the 3-month bill, and a policy shock of that size may act on a longer lag than past cycles did. The third is precedent. The curve inverted from September 1966 to January 1967, with a trough of −0.34 pt in October 1966, and USREC shows no recession start within 36 months; that episode sits outside the evaluation window, which starts with 1976 origins.

When will we know?

USREC is provisional at its end. It follows the NBER business-cycle dates, which are announced after the fact, so a peak dated later in 2025 or 2026 would turn recent zeros into ones. Two rules in this study take that into account. An episode is judged only once 36 months have passed since its onset and USREC covers that month. Until then it is pending, and it is never counted as an episode without a recession. A provisional cut also drops origins whose target falls in the last 18 USREC months, and on that cut the base model’s score is +0.103 (580 origins).

On those rules, the 30% episode that began in November 2022 is judged a false positive, because USREC covers its 36-month horizon of November 2025. The 50% episode that began in November 2023 (0.555 at −0.77 pt) is pending. Its 36-month horizon ends in November 2026, and it becomes judgeable once USREC covers that month.

History offers a second clock. In the inversions followed by a recession, the first recession month came 4 months after the last month of inversion in 2001, 9 months after it in 2008 and 6 months after it in 2020. In 1980 and 1981 it came 2 months before and in the same month. The 2022-24 run ended in November 2024. By October 2026, 23 months had passed with no USREC recession month, longer than any de-inversion lag in the table since 1953, where the longest is 9 months.

Neither clock is a forecast. They define when the 2022-24 record can be closed in the data, and as of November 2026 it cannot.

Slide the threshold, 1976 to 2026

The module below draws the base model’s probability since January 1976 over the USREC bands, with the known recession frequency for comparison and the spread in the tooltip. Move the threshold between 25% and 55% to see which episodes would have been flagged, and when. For the shape of the whole curve on any date, see the US Treasury yield curve through time.

The module reads the published CSV directly. If it does not load, the chart shown in its place is the static version of the default view.

What each threshold crossing was followed by

No market returns here. The study measures an indicator, so the outcome after each crossing is a recession start or its absence. An episode is a run of consecutive months at or above the threshold. It counts as a true positive if a recession starts within 24 months of its onset, and as a false positive if none starts within 36 months and USREC covers that horizon. Starts between 24 and 36 months stay unclassified, and episodes whose horizon USREC does not yet cover are pending.

At 30%, the base model flagged all 6 evaluated recessions with 8 true positives and 1 false positive, 2022-11. The median lead was 13.0 months, from 9 to 23. At 50% it flagged 2 of 6, in 1980 and 1981, with 5 true positives, 1 false positive (2022-12) and 1 pending episode (2023-11). Most 30% crossings before 2022 came while the spread was still positive, between +0.09 and +0.17 pt, whereas the 2022-11 crossing came at −0.26 pt.

Across variants, all five flag 6 of 6 recessions at 30%. At 50%, four of them flag 2 of 6, and the 20-year rolling window flags 3 of 6 by adding 2008.

Detections and false positives by threshold, 25% to 55%

Base model, evaluation origins from 1976. False positives are counted only when judgeable under the 36-month rule.

Source: FRED (GS10, TB3MS, USREC). Eco3min Research calculation on the published CSV.

The 30% line is a convention, and the sweep shows how much rides on it. From 25% to 30% all 6 recessions are flagged. From 31% the 1990 recession drops out, its peak reading having been 0.304 at +0.13 pt, then 2020 from 45%, 2008 from 47% and 2001 from 50%. Judgeable false positives number 2 at 25% and 26%, and 1 from 27% to 55%. Below 30%, some crossings fall 24 to 36 months before a recession and stay unclassified, 2 from 25% to 27% and 1 at 28% and 29%.

Past distributions are not predictive of future outcomes. Here the distribution is one of episode outcomes, built on 6 evaluated recessions, so one more episode can move every share in the table, and episodes at the same threshold overlap in time.

Levels to watch

Three markers, descriptive only, read from the October 2026 estimate. The coefficients are re-estimated each month, so the spread levels move with them.

Peak and decomposition tables

Two tables carry the episode detail behind the charts. The first gives the highest reading in the 24 months before each recession start, and the maxima over 36 months are the same. The second splits each inversion of 2 months or more into its long and short legs, from 24 months before the onset to the trough, and gives the lag from the last negative month to the recession start.

A negative lag means the recession began before the curve turned positive again. The 2020-03 recession is labelled an exogenous shock, the pandemic, and counted by protocol.

Historical turning points

Since April 1953 the spread has inverted in 13 runs, 11 of them lasting 2 months or more. The 2022-24 run ranks 1 of 13 on duration (25 months) and 3 on depth (−1.57 pt), behind 1980-81 (−2.65 pt) and 1978-80 (−2.45 pt). The 2s10s view of inversion history is in the yield-curve inversion history, and the dated record in the yield-curve inversion history dataset.

Runs of 2 months or more. Ranks are computed on all 13 runs, one-month runs included.

October 1966: inversion, no recession

The 1966-67 run lasted 5 months and bottomed at −0.34 pt in October 1966. USREC shows no recession start within the next 36 months, so the 2-month rule scores it a false positive. It sits before the evaluation window, which scores no forecast origin before January 1976.

December 1980: the deepest trough

The spread reached −2.65 pt in December 1980, and the model would have shown 0.992, the highest pre-recession reading of the 7 episodes. The recession started in August 1981, the same month the run ended.

February 2006: the longest lead

The base model crossed 30% in February 2006 at +0.14 pt, before the curve inverted in August 2006. The recession started in January 2008, 23 months later and 9 months after the last negative month.

August 2019: an exogenous shock

The 2019 run peaked at 0.446 at −0.32 pt in August 2019. The recession that started in March 2020 is an exogenous shock, the pandemic. The study counts it by protocol and reports the score without 2018-2020 origins, +0.073, alongside.

May 2023: the deepest since 1981

At −1.57 pt the 2023 trough was the lowest spread since May 1981 (−2.20 pt), and the model would have shown 0.766. The peaks ahead of the 1981 and 1980 recessions stood higher, at 0.992 in December 1980 (−2.65 pt) and 0.920 in December 1979 (−1.65 pt).

October 2026: the current observation

Methodology

Every probability in this study is a reconstruction: what a model re-estimated monthly on data then available would have shown, computed in November 2026. None of these readings was published at the time. “Real-time” refers to the estimation, which at each month uses only past spreads and recession dates lagged 12 months, a stand-in for when they became known. It does not refer to the choice of model, a probit on the 10-year/3-month spread published by Estrella and Hardouvelis in 1989-1991, first as a New York Fed research paper in May 1989.

Inference uses a circular block bootstrap with 24-month blocks, 2,000 draws and a fixed seed (20261022), for 90% intervals on the BSS, the pseudo-R² and the AUC. In the full sample, 1 draw contained no recession. The Diebold-Mariano test uses a HAC variance with 12 lags. The Clark-West test compares the spread model with the nested base-rate model. Both are one-sided.

Episodes follow a fixed algorithm. A run of consecutive months at or above the threshold is an episode, with its onset at the first month. The lead to a recession is measured from the onset of the first episode classified as a true positive for that recession. Pre-recession maxima are taken over 24 and 36 months, and the inversion decomposition runs from 24 months before the onset to the trough month.

Sensitivity covers the information lag (6 and 18 months), the rate convention (bond-equivalent), the estimation window (240 months), thresholds from 25% to 55% in steps of 1 point, maxima over 24 and 36 months, the sample without 2018-2020 origins, the provisional cut and the 24-month control. The cumulative edge stays positive in all five published variants, and the share of its end-2021 level erased by October 2025 runs from 50.2% (roll240) to 77.1% (lag6), with lag12 at 76.4%, lag18 at 77.1% (just below lag6 before rounding) and bond_equiv at 62.6%.

Recession periods follow the FRED series USREC, the Federal Reserve Bank of St. Louis reading of NBER business-cycle dates, where a month counts as recession from the period following the peak through the trough. No NBER date is used directly, and the NBER announcement delay is replaced by fixed lags of 6, 12 and 18 months. The probit on this spread goes back to Estrella and Hardouvelis (1991) and Estrella and Mishkin (1998). Estrella, Rodrigues and Schich (2003) tested its stability, and the forecast comparisons use Diebold and Mariano (1995) and Clark and West (2007).

The model columns are empty before January 1976 by construction. The code below rebuilds prob_rt_lag12 from the CSV, with a maximum gap of 5.00e-07 against the published column.

import numpy as np, pandas as pd, statsmodels.api as sm
from scipy.stats import norm
df = pd.read_csv("r23-real-time-yield-curve-recession-probability-1953-2026-v2.csv")
x, y, L = df["spread_10y3m"].values, df["usrec_t12"].values, 12
p = np.full(len(df), np.nan)
for i in df.index[df["date"] >= "1976-01"]:
    k = i - 12 - L                      # last origin s with s + 12 <= t - L
    m = sm.Probit(y[:k + 1], sm.add_constant(x[:k + 1])).fit(disp=0)
    p[i] = norm.cdf(m.params[0] + m.params[1] * x[i])
print(np.nanmax(np.abs(p - df["prob_rt_lag12"].values)))

Sources and references

• Board of Governors of the Federal Reserve System (US), Market Yield on U.S. Treasury Securities at 10-Year Constant Maturity, Quoted on an Investment Basis [GS10], retrieved from FRED, Federal Reserve Bank of St. Louis; fred.stlouisfed.org/series/GS10, November 3, 2026. Public domain, citation requested.
• Board of Governors of the Federal Reserve System (US), 3-Month Treasury Bill Secondary Market Rate, Discount Basis [TB3MS], retrieved from FRED, Federal Reserve Bank of St. Louis; fred.stlouisfed.org/series/TB3MS, November 3, 2026. Public domain, citation requested.
• Federal Reserve Bank of St. Louis, NBER based Recession Indicators for the United States from the Period following the Peak through the Trough [USREC], retrieved from FRED, Federal Reserve Bank of St. Louis; fred.stlouisfed.org/series/USREC, November 3, 2026. Copyrighted, citation required.
• Estrella, Arturo, and Gikas Hardouvelis (1991). “The Term Structure as a Predictor of Real Economic Activity.” Journal of Finance 46, 555–576. First circulated as a Federal Reserve Bank of New York research paper, May 1989.
• Estrella, Arturo, and Frederic S. Mishkin (1998). “Predicting U.S. Recessions: Financial Variables as Leading Indicators.” Review of Economics and Statistics 80, 45–61.
• Estrella, Arturo, Anthony P. Rodrigues and Sebastian Schich (2003). “How Stable Is the Predictive Power of the Yield Curve? Evidence from Germany and the United States.” Review of Economics and Statistics 85, 629–644.
• Diebold, Francis X., and Roberto S. Mariano (1995). “Comparing Predictive Accuracy.” Journal of Business & Economic Statistics 13(3), 253–263.
• Clark, Todd E., and Kenneth D. West (2007). “Approximately Normal Tests for Equal Predictive Accuracy in Nested Models.” Journal of Econometrics 138(1), 291–311.

Limitations

• Revisions. USREC follows NBER dates announced after the fact, so its last months are provisional. The provisional cut, which drops the last 18 months of targets, gives a base score of +0.103 instead of +0.054.
• Information lag. A fixed 12-month lag stands in for the date each recession became known. It approximates the announcement calendar without reproducing it, and the 6- and 18-month variants test that choice.
• Composition. Two yields and one binary flag. The model sees neither the term premium nor credit conditions, and a 10-year/2-year spread gives other dates, as the T10Y3M and T10Y2Y comparison shows.
• Non-stationarity. The link between the spread and recessions can shift with the monetary regime, the question Estrella, Rodrigues and Schich (2003) put to the US and German data. The 20-year rolling window is the variant built for that case, and it erases 50.2% of the edge.
• Overlapping windows. Monthly origins with 12-month targets overlap, so 598 origins hold 6 recessions and 58 recession target months. The 24-month block bootstrap and the HAC variance allow for that dependence, and they widen the intervals.
• Ex post against ex ante. The specification was published in 1989-1991. The 1976-1990 segment uses a model chosen with hindsight, which is why the 1991-2021 segment is reported beside it.
• Start date and scope. Scores begin with January 1976 origins, and every reading is a reconstruction made in November 2026. Market behaviour is outside the study, covered in why markets rose despite an inverted yield curve.

Frequently asked questions