Reverse-Engineering the NY Fed Recession Probability Model: How It Works, How Accurate It Is, and What It Says Now
Eco3min Research · Rates & Recession Indicators
Every U.S. recession since 1990 was preceded by a sustained 10Y–3M Treasury inversion. The more-cited 10Y–2Y produced a 32-month false positive in 1998.
The New York Fed publishes a monthly probability-of-recession estimate based on a single input: the spread between the 10-year Treasury yield and the 3-month Treasury bill rate (FRED series T10Y3M). comparing earnings yields and bond yields sets out the mechanism in detail. The model dates from Arturo Estrella and Frederic Mishkin’s 1996 paper and remains, three decades later, the most widely referenced quantitative recession indicator in U.S. financial media.
This study reconstructs the model on 35 years of daily Treasury data (1990–2024), replicates its probit specification with two parameter sets, compares the 10Y–3M signal head-to-head with the more popular 10Y–2Y spread, and documents the forward S&P 500 distributions associated with each yield-curve regime. The dataset covers 8 sustained inversion episodes of the 10Y–3M and 7 sustained episodes of the 10Y–2Y. The lead-time dispersion, the 1998 false-positive case, the unresolved 2022–2024 inversion, and the counterintuitive equity returns observed during deep inversions are documented in the sections below. All figures trace to the CSV linked at the end of this page (CC BY 4.0).
Latest Observation
As of 12 February 2026 (most recent FRED daily reading at time of analysis)
10Y–3M Spread
+0.39 pp
Positive; out of inversion since 2024-12-13
10Y–2Y Spread
+0.62 pp
Positive; out of inversion since 2024-08-26
Estrella & Trubin probit (12m)
29.7%
Implied recession probability, December 2024 reading (last with 12-month forward window data)
Months since 10Y–3M un-inversion
~14 months
Historical end-of-inversion-to-NBER-peak interval: 1 to 21 months (median ~5)
Sources: FRED series T10Y3M, T10Y2Y (daily readings at 12 February 2026); Estrella & Trubin (2006) parameters applied to December 2024 monthly average.
Executive Summary
- Every U.S. recession since 1990 was preceded by a sustained 10Y–3M Treasury inversion. The more-cited 10Y–2Y produced a 32-month false positive in 1998.
- Over 1990–2024 the 10Y–3M spread recorded 8 sustained inversion episodes: 6 true positives covering the three in-window NBER recessions of 2001, 2007–09 and 2020 (multiple inversions preceded the 2007–09 and 2020 recessions), 1 unresolved (2022–2024), and 1 coincident (brief Feb 2020 inversion overlapping the COVID recession start). Zero false positives.
- The lead time from 10Y–3M inversion onset to NBER recession peak ranged from 4.4 to 21.3 months across the six true positives (median 9.4 months). This dispersion is too wide for precise market timing but consistent enough to function as a regime indicator.
- The Eco3min replication of the Estrella probit (1990–2024, n=408 monthly observations with a 12-month forward window) yields α = −0.85, β = −0.50, pseudo-R² = 15.6%. Applied to May 2023 data, the model produced a 50.4% implied probability of recession within 12 months — the highest reading on record under our parameters; the original Estrella & Trubin (2006) parameters yielded 69.3% at the same observation.
- S&P 500 forward 12-month returns conditional on the prevailing T10Y3M regime are distributed as follows: Strong Inverted (<−1pp, n=11) median +25.3% with 100% positive observations; Mild Inverted (−1 to 0, n=26) median +4.5% with 53.8% positive and median MDD −14.6%; Flat (0 to +1pp, n=97) median +13.8% with 81.4% positive; Normal Steep (>+1pp, n=274) median +10.5% with 79.9% positive. Past distributions are not predictive of future outcomes.
Stats Panel
3 / 3
In-window NBER recessions (2001, 2007–09, 2020) preceded by a sustained 10Y–3M inversion
1
False positive in the 10Y–2Y series (1998, 32.5-month gap)
9.4 m
Median lead time, 10Y–3M onset to NBER peak (range: 4.4–21.3)
−1.89 pp
Deepest 10Y–3M inversion on record (2023-05-04, 2022–24 episode)
69.3%
Peak Estrella & Trubin probit reading (May 2023)
11.8%
Share of daily observations 1990–2024 with negative T10Y3M
1. How the NY Fed Recession Probability Model Works
The dominant mainstream framing treats yield-curve inversion as a binary recession signal: when the curve inverts, a recession follows; when it does not, no recession is expected. This binary reading appears in market commentary, central-bank speeches, and financial media headlines. The NY Fed’s published probability series — derived from a probit regression on the monthly average T10Y3M spread — refines this into a continuous probability between zero and one. The full methodological breakdown of this probit regression — parameters, calibration window, robustness across samples — is documented in the full breakdown of the NY Fed probit model.
The Estrella & Mishkin (1996) specification regresses the binary indicator recession-within-12-months on the monthly average 10Y–3M spread, using the cumulative standard normal distribution:
P(recession in next 12 months) = Φ(α + β · spreadt)
Where Φ is the standard normal CDF, α is the intercept, β is the spread coefficient (expected to be negative), and spreadt is the monthly average T10Y3M in percentage points. The NY Fed’s published version updates the parameters periodically using the 1959–present sample. The most recent fully documented parameter set (Estrella & Trubin, 2006) reports α = −0.5333 and β = −0.5984 estimated on 1959–2009 data.
Replicating the specification on Eco3min’s 1990–2024 sample (n = 408 monthly observations with a non-missing 12-month forward recession indicator, 34 of which fall in the 12-month window preceding an NBER recession) produces α = −0.8493, β = −0.4954, with both coefficients statistically significant at p < 0.001 and a McFadden pseudo-R² of 0.156. The lower intercept in the Eco3min fit reflects the lower in-sample base rate of recession in the 1990–2024 sample (34/408 = 8.3%) relative to 1959–2009. Applied to the May 2023 observation (monthly average T10Y3M = −1.73 pp), the Estrella & Trubin parameters yield 69.3% probability; the Eco3min parameters yield 50.4%. Both readings are the highest in their respective series and exceed the highest pre-recession ET reading recorded in advance of the 2007–09 GFC (41.1% in March 2007) by approximately 28 percentage points.
This is a one-input model. It does not incorporate credit spreads, real activity, labor-market conditions, or financial-condition indices. Its appeal lies in parsimony — and in the empirical observation that a single Treasury spread has performed comparably to richer multivariate models across most of the post-WWII sample. The empirical record over 1990–2024 is examined in the next section.
2. The 1998 Divergence: Where the Two Spreads Disagreed
The most direct empirical test of which yield-curve measure performs better as a recession signal is the 1998 episode. Between 15 June and 27 July 1998 the 10Y–2Y spread inverted on a sustained basis for 23 trading days, reaching a trough of −0.07 pp. The trigger was the Russia debt default in August 1998 and the subsequent collapse of Long-Term Capital Management (LTCM) in September, which compressed long-end yields through a flight-to-quality bid. Over the same period the 10Y–3M spread also flattened but did not sustain a negative reading. Across all of 1998, the 10Y–3M printed negative on only 5 trading days (none consecutive), reaching its 1998 minimum of −0.13 pp on 21 September.
The next U.S. recession began in March 2001 — 32.5 months after the onset of the 1998 10Y–2Y inversion. By any conventional 24-month threshold this was a false-positive signal from the 10Y–2Y. The 10Y–3M, which never sustained inversion in 1998, did invert in July 2000 (sustained, 133 trading days, trough −0.95 pp), producing a 7.8-month lead to the March 2001 recession — within the normal range. On this single but high-profile episode, the 10Y–3M produced no signal while the 10Y–2Y produced a false signal, and the recession arrived on a schedule consistent with the 10Y–3M’s later inversion. The 1998 episode is documented in greater historical detail in the Eco3min 2s10s history page.
Across the full 1990–2024 sample, the 10Y–2Y produced 7 sustained inversion episodes (5 true positives, 1 false positive, 1 unresolved 2022–2024). The 10Y–3M produced 8 episodes (6 true positives, 0 false positives, 1 coincident-with-recession brief February 2020 inversion, 1 unresolved 2022–2024). The 10Y–3M’s superior precision on this sample is consistent with Estrella & Mishkin’s (1996) original finding that the 10Y–3M provides the cleanest signal of the standard term-spread family.
A legitimate analytical qualification: the 1998 false positive sample is one observation. Generalizing from a single episode to the conclusion that the 10Y–2Y is structurally less reliable is statistically thin. The Estrella & Mishkin sample including pre-1990 data identifies a 1966 episode that they classify as a false positive for both spreads. The 1989 10Y–3M inversion preceded the July 1990 recession with a lead time documented in academic sources (Estrella & Mishkin, 1996, Table 1) but falls outside the daily-resolution window available from the U.S. Treasury archive on which this study is based. With these qualifications acknowledged, the 1990–2024 record nonetheless gives the 10Y–3M one fewer false positive in the same window — and the unresolved 2022 inversion casts that comparison forward.
3. Lead-Time Dispersion Across Recessions
The “median 12-month lead time” frequently cited in financial commentary obscures wide dispersion. Across the six true-positive 10Y–3M episodes between 1990 and 2024, the lead time from sustained-inversion onset to NBER business-cycle peak ranged from 4.4 to 21.3 months. The values, in chronological order: 7.8 months (July 2000 → March 2001), 21.3 months (February 2006 first brief inversion → December 2007), 16.4 months (July 2006 sustained inversion → December 2007), 4.4 months (July 2007 brief reinversion → December 2007), 10.4 months (March 2019 first inversion → February 2020), and 8.3 months (May 2019 sustained inversion → February 2020).
This dispersion has two structural sources. First, the 2006–2007 cycle produced multiple inversions of differing duration before the GFC recession, depending on which onset date is selected. Second, the COVID recession of 2020 was triggered by an exogenous shock unrelated to the credit-cycle mechanism through which yield-curve inversion is generally understood to operate. The 8.3-month lead in the 2019 episode therefore deserves a steelman qualification: had no pandemic shock occurred in early 2020, it is unclear whether the May 2019 inversion would have been followed by a recession on a similar timeline. The transmission channel — bank net-interest-margin compression and the resulting tightening of credit standards — typically requires 12–18 months to feed through to aggregate demand. Some researchers therefore treat the 2019 inversion as correlation with, not cause of, the 2020 recession.
The same dispersion appears in the 10Y–2Y series. For the 1998 false positive specifically, the gap to the next recession was 32.5 months — well outside any defensible lead-time window — which is why this episode is classified as a false positive rather than a slow-burn true positive. A 24-month cutoff is the standard convention in the academic literature on this question (see Bauer & Mertens, 2018).
The practical implication is that the inversion signal is a regime indicator (recession-risk environment is structurally elevated) rather than a precise timing instrument. The lead-time data over 1990–2024 do not support claims of consistent 12-month forecasting precision.
4. Forward Returns by Yield-Curve Regime
S&P 500 forward 12-month returns conditional on the prevailing T10Y3M regime show a non-monotonic pattern. The four regimes are defined as: Strong Inverted (spread < −1 pp), Mild Inverted (−1 to 0 pp), Flat (0 to +1 pp), and Normal Steep (> +1 pp). Across 408 monthly observations with full forward windows (1990-01 through 2023-12):
| Regime (T10Y3M) | n (months) | Median 12m return | IQR (P25–P75) | % positive | Median 12m MDD | Worst 12m MDD |
|---|---|---|---|---|---|---|
| Strong Inverted (<−1 pp) | 11 | +25.3% | +19.7% to +29.1% | 100% | −8.6% | −8.6% |
| Mild Inverted (−1 to 0 pp) | 26 | +4.5% | −9.6% to +17.0% | 53.8% | −14.6% | −27.5% |
| Flat (0 to +1 pp) | 97 | +13.8% | +5.6% to +22.1% | 81.4% | −8.6% | −41.0% |
| Normal Steep (>+1 pp) | 274 | +10.5% | +2.0% to +21.9% | 79.9% | −5.8% | −47.5% |
Past distributions are not predictive of future outcomes. Regime-conditional statistics describe historical patterns, not expected returns. Sample size for the Strong Inverted bucket (n=11) is concentrated almost entirely in the 2022–2024 episode, during which the equity market posted exceptional returns; this is a small-sample warning, not a generalizable pattern. MDD = maximum drawdown over the forward 12-month window.
Two patterns stand out. First, the Mild Inverted regime — not the deepest inversions — shows the lowest median forward return and the highest probability of negative outcomes. Of the 26 monthly observations in this bucket, less than 54% delivered a positive 12-month return, and the median maximum drawdown over the forward window was −14.6%. This regime contains most of the 2000–2001 and 2007 pre-recession observations.
Second, the Strong Inverted bucket is dominated by the 2022–2024 episode, during which the S&P 500 posted consecutive years of strong gains while the curve was at record-depth inversion. This produces a counterintuitive headline median of +25.3% over 12 months. The bucket is small (n = 11), concentrated in a single episode, and reflects the unresolved status of the 2022 inversion. It does not establish a regularity that deep inversions are bullish. The pattern is consistent with two complementary readings: equity prices tend to peak during or after — not before — inversion onset, and the 2022–2024 episode may yet prove to be a delayed-recession case or a structural false positive. The forthcoming forward-12-month returns dated from observations in 2024 will further inform this distinction as the data window closes.
Surveillance: Levels to Watch
As of the most recent FRED reading (February 2026), the 10Y–3M spread sits at +0.36 pp, having un-inverted in mid-December 2024. The following observables would alter the current regime classification:
Invalidation threshold
If the monthly average 10Y–3M spread closes below 0 pp for two consecutive months, the regime would re-enter Mild Inverted territory. Under the Estrella & Trubin parameters, a monthly average of −0.50 pp would produce an implied 12-month recession probability of approximately 41%; a reading at −1.00 pp would produce approximately 53%.
Confirmation signal
If the 10Y–3M widens above +1.0 pp on a sustained monthly basis, the regime would shift to Normal Steep, which historically (1990–2024, n = 274 monthly observations) has been associated with an 80% probability of positive forward 12-month S&P 500 returns and a median maximum drawdown of −5.8% over the forward window. This is a historical observation, not a forecast.
Calendar catalyst
The U.S. Treasury publishes T10Y3M via FRED on each business day. The NBER Business Cycle Dating Committee meets without a fixed schedule; recession determinations historically lag the business-cycle peak by 6–18 months. The unresolved 2022–2024 inversion would be re-classified as either a true-positive or a definitive false-positive depending on whether NBER dates a recession between the inversion start (October 2022) and a window cutoff that researchers conventionally set at 36 months post-onset — meaning the case becomes final no later than October 2025 under the most lenient academic threshold. As of the time of writing, no NBER peak has been announced.
Historical Turning Points (1998–2024)
Side-by-side detail of the five episodes most cited in commentary on yield-curve recession signals. The 10Y–3M and 10Y–2Y dates are independent: the same macroeconomic episode may produce different signal characteristics on the two spreads.
| Episode | 10Y–3M onset | 10Y–2Y onset | 10Y–3M trough | 10Y–2Y trough | Next recession | 10Y–3M verdict | 10Y–2Y verdict |
|---|---|---|---|---|---|---|---|
| 1998 Russia/LTCM | No sustained inversion | 1998-06-15 | −0.13 pp (5 brief days) | −0.07 pp | 2001-03 (32.5 months later) | No signal | False Positive |
| 2000–01 Dot-com | 2000-07-07 | 2000-02-02 | −0.95 pp | −0.52 pp | 2001-03 | True Positive (7.8m lead) | True Positive (12.9m lead) |
| 2006–07 GFC | 2006-07-19 (sustained) | 2006-01-31 | −0.64 pp | −0.16 pp | 2007-12 | True Positive (16.4m lead) | True Positive (22.0m lead) |
| 2019 COVID-adjacent | 2019-05-23 (sustained) | 2019-08-27 | −0.52 pp | −0.04 pp | 2020-02 | True Positive (8.3m lead) * | True Positive (5.2m lead) * |
| 2022–24 Post-COVID | 2022-10-25 | 2022-07-06 | −1.89 pp (2023-05-04) | −1.08 pp | None as of 2024-12-31 | Unresolved (30+ months) | Unresolved (29+ months) |
* The 2020 recession was triggered by an exogenous shock (COVID-19 pandemic). The causal link to the preceding 2019 inversion remains debated in the academic literature. Sustained inversion = ≥2 consecutive months with negative monthly average spread. Trough dates and depths from daily T10Y3M / T10Y2Y series.
Methodology & Sources
Data sources
Treasury yields. Daily constant-maturity yields for the 3-month bill, 2-year note, and 10-year note from 1990-01-02 to 2024-12-31 are sourced from the U.S. Department of the Treasury (Daily Treasury Par Yield Curve Rates archive). This is the underlying series for FRED’s DGS3MO, DGS2, and DGS10. Spreads are computed by subtraction.
Recession dates. National Bureau of Economic Research, Business Cycle Dating Committee. The four NBER-dated U.S. recessions within the 1990–2024 window: July 1990 to March 1991; March 2001 to November 2001; December 2007 to June 2009; February 2020 to April 2020.
Equity prices. S&P 500 daily closes from Yahoo Finance (ticker ^GSPC) over the same window, used to compute forward 6-month and 12-month percentage returns and forward 12-month maximum drawdowns.
Probit specification
The probit regression follows the Estrella & Mishkin (1996) form: P(recession in next 12 months) = Φ(α + β · spreadt), where Φ is the standard normal CDF and spreadt is the monthly average T10Y3M in percentage points. The dependent variable is 1 if any month within the next 12 months is in an NBER-dated recession, 0 otherwise. Estimation is by maximum likelihood on 408 monthly observations (January 1990 to December 2023; the last 12 months of the daily window are dropped because the 12-month-ahead recession indicator is undefined when the lead window extends beyond the data sample). Eco3min parameters: α = −0.8493 (s.e. 0.12, p < 0.001), β = −0.4954 (s.e. 0.09, p < 0.001), McFadden pseudo-R² = 0.156. The Estrella & Trubin (2006) parameters (α = −0.5333, β = −0.5984) are quoted from their published Table 1 and applied without re-estimation, as a comparison point.
Inversion episode definition
A sustained inversion episode is defined as the first day on which the daily spread turns negative and remains negative on a 5-trading-day rolling-average basis for at least 5 consecutive trading days. The episode ends on the last day before the spread returns durably to positive territory under the same criterion. Brief intraday or single-day inversions are excluded — a convention consistent with the NY Fed’s monthly-average reporting framework. Episode-level signal classification uses a 24-month forward window: an inversion preceding an NBER peak within 24 months is a True Positive; an inversion not followed by an NBER peak within 36 months is a False Positive; an inversion within the past 36 months without an NBER peak is Unresolved. The brief February 2020 10Y–3M inversion is classified Coincident as it occurred during the NBER-dated recession itself.
Pre-1990 data and the 1989 inversion
The U.S. Treasury’s daily Par Yield Curve Rates archive begins in 1990. The 10Y–3M inversion that preceded the July 1990 recession by approximately 9 months is documented in Estrella & Mishkin (1996, Table 1) using a separate FRED daily series that begins in 1982. That inversion is not visible in this study’s daily dataset because the spread had already un-inverted by January 1990. The 1990 recession that begins at the start of the sample is therefore not preceded by an inversion within the daily window analyzed here, though it was preceded by one in the broader historical record. The episode count for 1990–2024 in this study excludes the pre-1990 inversion.
Limitations
Two structural limitations are worth noting. First, the 1990–2024 daily window covers only four NBER recessions. The lead-time dispersion and false-positive base rate estimates rely on small samples. Confidence intervals on these estimates are wide. Second, the equity forward-return distributions in the Strong Inverted bucket are concentrated in the 2022–2024 episode (n = 11 observations). Generalizing from this cluster to a broader regime characterization would be statistically unreliable.
Academic references
- Estrella, A. and Mishkin, F. (1996). “The Yield Curve as a Predictor of U.S. Recessions.” Federal Reserve Bank of New York, Current Issues in Economics and Finance, Vol. 2, No. 7.
- Estrella, A. and Trubin, M. (2006). “The Yield Curve as a Leading Indicator: Some Practical Issues.” Federal Reserve Bank of New York, Current Issues in Economics and Finance, Vol. 12, No. 5.
- Engstrom, E. and Sharpe, S. (2018, 2022). “(Don’t Fear) The Yield Curve.” FEDS Notes, Board of Governors of the Federal Reserve System.
- Bauer, M. and Mertens, T. (2018). “Information in the Yield Curve about Future Recessions.” FRBSF Economic Letter, No. 2018-20.
- Harvey, C. (1986). Recovering Expectations of Consumption Growth from an Equilibrium Model of the Term Structure of Interest Rates. Doctoral dissertation, University of Chicago.
Frequently Asked Questions
Why does the NY Fed use the 10Y–3M spread rather than the more popular 10Y–2Y?
The 10Y–3M anchors the short rate to current Federal Reserve policy, since the 3-month bill yield tracks the federal funds rate closely. The 10Y–2Y embeds the market’s forward path of Fed policy 12 to 24 months out, which means it can invert on changes in expectations even when current policy is not particularly restrictive. Estrella & Mishkin’s 1996 paper found the 10Y–3M produced the strongest in-sample fit among the term spreads they tested, and the NY Fed has used it as the base of its published probability series ever since. On the 1990–2024 sample documented here, the 10Y–3M produced zero false positives compared with one (the 1998 episode) for the 10Y–2Y.
What does the model say right now?
As of the most recent FRED reading at the time of analysis (12 February 2026), the 10Y–3M spread sits at +0.39 pp, having un-inverted in mid-December 2024 after 534 consecutive trading days in negative territory. The Estrella & Trubin parameters applied to the December 2024 monthly average yield an implied 12-month recession probability of 29.7%, well below the May 2023 peak of 69.3%. This reading is similar to levels recorded in late 1995 and mid-2007, both of which preceded recessions by 12+ months — but also similar to readings in 2003 and 2014, which did not. The current snapshot is below the model’s pre-recession threshold but above the long-run average of approximately 13%.
Is the 2022–2024 inversion a false positive?
The case remains unresolved at the time of analysis. The 10Y–3M was inverted from 25 October 2022 through 12 December 2024 (534 trading days, deepest at −1.89 pp on 4 May 2023). The 10Y–2Y was inverted from 6 July 2022 through 26 August 2024 (537 trading days, deepest at −1.08 pp). No NBER-dated recession has begun as of the data cutoff. By a strict 24-month forward window, both signals would be classified as false positives. By the 36-month window used in some academic studies, the 10Y–3M episode would remain valid through October 2025 and the 10Y–2Y through July 2025. Eco3min classifies both as Unresolved pending further data and any future NBER announcement.
How accurate is the NY Fed probit model in practice?
On the 1990–2024 sample, the Eco3min replication produces a McFadden pseudo-R² of 0.156. This is a modest fit by general statistical standards, but it is achieved with a single input variable. The Estrella & Trubin probit reading exceeded 30% within the 12 months preceding each of the three NBER recessions for which an in-window 10Y–3M signal is available: peak 45.4% in the run-up to the 2001 recession, 41.1% in March 2007 ahead of the GFC, and 37.6% in August 2019 ahead of the COVID recession. It also exceeded these thresholds during the 2022–2024 reading without (so far) being followed by a recession. The model’s directional record over 1990–2024 is consistent — but the implied probability magnitudes have varied substantially from cycle to cycle, and the May 2023 reading of 69.3% has no precedent that was followed by a recession in this sample.
Why do equities tend to perform well during inversions?
The forward-return data show a counterintuitive pattern: the Strong Inverted bucket (spread < −1 pp) had a median 12-month forward S&P 500 return of +25.3% across 11 observations. This is concentrated almost entirely in the 2022–2024 episode and does not generalize. Two complementary mechanisms apply across the broader sample. First, equity markets typically peak during or after inversion onset, not before — meaning the period of deepest inversion often coincides with continued price appreciation that ends only as the recession itself begins. Second, deep inversions can compress long-term yields, supporting equity valuations through the discount-rate channel. The more concerning bucket statistically is Mild Inverted (−1 to 0 pp), where the median forward 12-month return falls to +4.5% with only 53.8% of observations positive, and median forward maximum drawdown reaches −14.6%.
How does the un-inversion relate to recession timing?
An empirical regularity often cited is that recessions tend to begin after the curve has re-steepened from inversion, not during the inversion itself. In the 1990–2024 sample, the gap from the end of a sustained 10Y–3M inversion episode to the next NBER business-cycle peak ranged from 1 to 21 months, with a median of approximately 5 months (6 true-positive episodes: 2000, 2006 brief, 2006 sustained, 2007, 2019 brief, 2019 sustained). The mechanism: un-inversion typically reflects short-term rates falling faster than long-term rates as the Fed begins easing, which usually coincides with deteriorating economic data. The most recent 10Y–3M un-inversion occurred on 13 December 2024, placing the calendar window during which a recession would historically have begun in the December 2024 – September 2026 range — a window that, at the time of writing in February 2026, has not produced an NBER-dated peak.
How does this model compare with the near-term forward spread?
Engstrom & Sharpe (2018, 2022) propose an alternative forward spread that isolates monetary-policy expectations over a 6-quarter horizon (essentially the difference between the 6-quarter-ahead 3-month rate implied by Treasury yields and the current 3-month rate). They find that once this near-term forward spread is included in a probit specification, the 10Y–3M coefficient becomes statistically insignificant. The implication is that the predictive content of the long-vs-short spread reduces to the short-end forward path of Fed policy. Some Fed researchers therefore consider the near-term forward spread to dominate the traditional measure. The 10Y–3M remains the more widely tracked indicator in market commentary, partly due to its longer track record and the NY Fed’s continued publication of the probit series in this form.
Dataset Download
The full daily and monthly dataset is available under CC BY 4.0. Use with attribution: Eco3min, NY Fed Recession Probability Model dataset (2026).
- Monthly dataset — Treasury spreads, probit probabilities, NBER flags, S&P 500 forward returns (420 rows, CSV, 40 KB)
- Inversion episodes table — 15 sustained episodes 1990–2024 with signal classification (CSV, 2 KB)
The monthly file contains columns for 3-month, 2-year, and 10-year Treasury yields, both spread measures, inversion flags, NBER recession dummy, both probit probability series, S&P 500 closes, forward returns and drawdowns, and regime classification. Methodology and code are documented in the section above.
Conclusion
The NY Fed’s 10Y–3M recession probability model has produced six true positives, one coincident signal, and one unresolved case across the 1990–2024 daily sample documented in this study, with zero clean false positives. Its closest cousin, the 10Y–2Y spread cited more frequently in financial media, has one documented false positive (1998 Russia/LTCM, 32.5 months before the next recession) and one unresolved case (2022–2024) over the same period. The Eco3min replication of the probit specification yields a peak implied probability of 50.4% in May 2023 (69.3% under the original Estrella & Trubin parameters) — both readings exceed any prior peak in the series.
The 2022–2024 episode — the longest and deepest 10Y–3M inversion on record — has yet to be followed by an NBER-dated recession at the time of writing. This is the central open question in the literature. Until that case resolves, the empirical record of the model spans six clear true positives, one false positive on the alternative 10Y–2Y measure, and one period whose classification remains in suspension. The data and replication code are released under CC BY 4.0 for further analysis.
The data and analysis presented on this page are provided for informational and educational purposes only. They do not constitute investment advice or a recommendation to take any specific action.
Last updated — 23 May 2026
Disclaimer – Financial Information: The analyses, commentary, and content published on eco3min.fr are provided for informational and educational purposes only. They do not constitute investment advice or a solicitation to buy or sell financial instruments. Past performance is not indicative of future results. All investment decisions involve risk and are the sole responsibility of the reader.