The Federal Reserve Bank of New York’s recession model now puts the probability of a US recession in September 2027 at 12.76%, down sharply from the readings above 70% it produced during the yield curve inversion of 2022-2024. That earlier alarm was wrong: no recession followed.
The model rests on the US Treasury yield curve, the most celebrated recession indicator in economics. An inversion preceded eight of the nine US recessions recorded between 1960 and 2020, on a rule that requires the signal to persist and a forecast horizon of 24 months.
The record looks extraordinary at first glance. It hides real statistical weaknesses. Flagging the arrival of a downturn is one thing; attaching trustworthy probabilities to each alarm is another.
Forecasting recessions has always been the most ambitious task in economic analysis, because a contraction means more unemployment, falling corporate profits, and markets that change their mood. Decades of study and ever more sophisticated econometrics have not made pinpointing the arrival of a downturn much easier. Some recessions build for years beneath the surface; others erupt like a summer thunderstorm.
A recession is only identified after the fact
Even the definition is not obvious. The most widely used convention, two consecutive quarters of falling gross domestic product, does not match the method followed in the United States.
The National Bureau of Economic Research, which officially establishes the chronology of US business cycles, looks at the overall performance of economic activity. Its indicators include employment, real household income, industrial production, consumer spending, and business sales.
A recession is identified when the deterioration is broad and widespread enough. Duration matters, but a particularly violent contraction can qualify even if it lasts only a few months, as in 2020.
For anyone making economic and financial decisions, the method has a drawback. Recessions are generally recognized once they have already begun, sometimes many months later. In December 2008, the NBER announced that the US economy had reached its cyclical peak in December 2007.
Learning a year late that you are in a recession is of little use to investors and businesses. Hence the search for indicators that can get ahead of the cycle.
The yield curve and the mechanism that anticipates downturns
The yield curve relates bond interest rates to their maturities. In normal conditions, longer-dated securities yield more than short-dated ones, rather like a term deposit that pays more for every year you cannot touch your capital. Investors demand extra compensation for tying up their money and for bearing the risks attached to inflation and the path of interest rates.
When short-term yields exceed long-term ones, the curve inverts. That anomalous configuration can reflect restrictive monetary policy and an expectation that the central bank will be forced to cut rates later, perhaps to counter a weakening economy.
Economists Arturo Estrella and Mary R. Trubin examined the link in depth in research published in 2006 by the Federal Reserve Bank of New York, The Yield Curve as a Leading Indicator: Some Practical Issues.
They identify the spread between the yield on the 10-year Treasury and that on the 3-month Treasury as one of the most effective indicators. When the difference turns negative, the probability of a subsequent recession historically rises.
The choice of maturities matters. The 10-year and 2-year spread, which markets watch frequently, tends to invert earlier and more often, while the 10-year and 3-month combination has held up better in historical testing. So does the length of the signal: a few days of inversion can reflect passing moves in the bond market, while an inversion that holds for months normally carries more information.
Eight recessions out of nine across nearly seventy years of data
Test the indicator on a monthly series running from January 1959 to September 2026 and you need a rule that screens out sporadic inversions. Take this one: the alarm counts as active when the spread is negative for at least three months within six consecutive months, and the signal counts as correct when a recession begins within the following 24 months. On that basis, the yield curve anticipated eight of the nine recessions recorded in the United States between 1960 and 2020.
The signal came eleven months before the 1970 downturn and four months before the 1973 one. For the 1980 and 1981 recessions the warning arrived thirteen and eight months ahead. In 1990 the lead time reached twelve months, in 2001 it fell to seven, and before the Great Recession that began in 2008 it stretched to fifteen. That works out to an average lead time of about 9.6 months across the eight correct signals. The only recession in the period not preceded by a persistent inversion is the one in 1960.
A curve inversion in the summer of 2019 also preceded the 2020 recession. But that contraction was caused by the pandemic emergency, an event whose specific features the 2019 bond market could not have foreseen, and the coincidence in timing alone does not prove the curve had sensed its causes.
The record also includes two significant false alarms, in 1966-1967 and in 2022-2024. Out of ten distinct alarm episodes, eight were followed by a recession within 24 months, a success rate of 80%. The result is notable, but it rests on too few downturns to carry high statistical precision, and it depends on the minimum inversion length and the 24-month horizon chosen to judge the signal correct.
The Fed model’s probabilities put to the test
The Federal Reserve Bank of New York uses a statistical model known as a probit, which converts the bond spread into a recession probability.
The forecast concerns a specific condition: the probability that the economy will be in recession exactly twelve months after the yields are measured. That is not the same as the probability that a recession will begin at some point during the following year.
Across 801 monthly observations from January 1960 to September 2026, where the predicted probability can be compared with the economy’s actual performance, the model achieves an AUC of 0.831.
AUC measures the ability to distinguish periods that will prove recessionary from those that will not. A value of 0.5 means discriminating power no better than chance, while 1 indicates perfect classification.
At 0.831, the model separates phases of differing risk well. Reading that number as an 83.1% rate of correct forecasts would be a mistake.
Measuring the reliability of the alarms requires a threshold. Choose 30%, a particularly meaningful level because it corresponds roughly to the point at which the curve inverts, and the results look less encouraging.
Across the full sample, the model produced 120 monthly forecasts at or above 30%. Of these, 50 were associated with months that did turn out to be recessionary twelve months after the signal was recorded. Seventy were not.
Only 41.7% of forecasts above 30% therefore translated into a recessionary month at the predicted horizon. The model also captured only 52.6% of the recessionary months actually observed at this threshold.
The contrast with the 80% success rate across historical episodes is only apparent, because the two statistics measure different events. An indicator can correctly flag the onset of a recession while missing the exact month it shows up, and a forecast that stays elevated for many months generates many statistical observations even though it concerns a single economic cycle. For anyone steering financial decisions, detecting an increase in recession risk is not the same as knowing when the recession will start.
The 2022-2024 false alarm changes the assessment of the indicator
The US 10-year/3-month curve stayed inverted for 25 consecutive months, from November 2022 to November 2024. In May 2023 the spread reached roughly -1.71 percentage points, one of the deepest negative readings of recent decades.
The probit model put the recession probability for May 2024 at 70.85%. The recession did not happen. The US economy kept expanding, and as of October 27, 2026, the most recent recession officially recognized by the NBER remains the one that ended in April 2020.
The monthly probabilities make the false alarm starker still. For 25 consecutive months of forecast reference, from November 2023 to November 2025, the model indicated a recession probability at or above 30% without a single recessionary month being recorded.
The problem goes beyond this episode. Split the data used to estimate the model’s parameters, which run through December 2009, from the data that follow, and the reliability of the probabilities deteriorates sharply.
From 2010 to September 2026, the model assigned an average recession probability of 15%. The actual frequency of recessionary months was about 1%. All 32 monthly forecasts above the 30% threshold turned out to be false positives against the goal of identifying a recessionary month exactly twelve months after the signal.
The post-2009 period contains only two recessionary months, both tied to the pandemic. Such an unbalanced sample calls for caution, since it cannot settle whether the model has structurally lost predictive power. The probabilities derived from the historical relationship between rates and recessions have nonetheless proved poorly calibrated in recent years.
Other indicators to use alongside the curve
The yield curve describes investor expectations but says nothing about the current state of the economy. Pairing it with indicators of a different kind gives a fuller picture of recession risk.
One of the best known is the Sahm Rule, developed by economist Claudia Sahm. The signal triggers when the three-month average of the unemployment rate exceeds the lowest level that same average reached in the previous twelve months by at least half a percentage point.
The principle is tied to labor market dynamics. When unemployment starts rising rapidly, the resulting drop in incomes and consumer spending can deepen the slowdown.
Unlike the yield curve, the Sahm Rule usually serves to recognize a deterioration already under way rather than to anticipate a recession a year ahead. It too can give false alarms.
A second tool is the Leading Economic Index, compiled by the Conference Board. The index combines several leading variables, among them labor market indicators, new orders, financial conditions, and economic expectations.
The latest reading shows the US LEI unchanged in September 2026, after a slight decline in August. In the six months ending in September, the index posted growth of 0.6%, reversing the contraction of the previous half-year.
Spreads between corporate bonds and government securities serve as a warning light too. When they widen, they can signal rising perceived risk and worsening corporate financing conditions. Industrial production, manufacturing orders, and consumer spending show whether the weakening in financial expectations is reaching the real economy.
The methodological knot is how to combine these signals without giving them equal weight, since some anticipate turning points and others react once the slowdown is under way. A model that uses them together could reduce false alarms, provided its superiority is demonstrated with genuinely out-of-sample testing.
What the Fed model says today about the outlook for the US economy
In September 2026 the spread between the 10-year Treasury yield and the 3-month bill reached +0.955 percentage points, or roughly 95.5 basis points. The curve has regained a positive slope after the long inversion that began in 2022.
Applying the New York Fed model, the probability that the US economy will be in recession in September 2027 comes out at 12.76%.
That figure sits well below the peaks of the previous inversion episode and points to a scaling back of the recession risk implied by the structure of yields. Read it with caution. It concerns only the historical relationship between the bond spread and economic activity; it is neither a comprehensive forecast combining all available macroeconomic information nor an official US monetary policy projection, and it comes from a model with recent calibration problems. There is not enough evidence to claim that the actual probability of a recession is precisely the one indicated. Treat the number as a relative measure: the risk the curve is signaling today is appreciably lower than what was observed during 2023.
Why no single indicator can predict every recession
Recessions do not all have the same origin. Some follow restrictive monetary policy, investment excesses, or credit imbalances. Others are born of financial crises, energy shocks, geopolitical tensions, or sudden events. An indicator built on past relationships can recognize imbalances already priced into markets with some effectiveness, but it is poorly equipped against external events that abruptly change how the economy works.
The time horizon matters as well. A signal can be informative over a twelve- or twenty-four-month horizon and nearly useless when you try to pinpoint the exact quarter in which the contraction will begin.
Eight recessions anticipated out of nine is a historically significant result. The 2022-2024 false alarm shows how risky it is to mistake a statistical correlation for an immutable relationship, because an indicator with an impeccable reputation can produce profoundly wrong forecasts in a single economic phase, like a clock that kept perfect time for decades and then stopped without warning.
Market watchers can use recession probabilities to frame the economic outlook. They should not automatically turn them into buy or sell signals.
The yield curve remains a useful tool for gauging strains in the business cycle. But it resembles a compass calibrated for terrain that has since changed, and until we know how far to trust the needle when conditions shift, recession forecasts will remain an exercise with limits that are hard to overcome.
Editor’s note
This article was originally published in Italian on money.it by Gerardo Marciano on October 31, 2026 as «Recessione USA, ha previsto 8 crisi su 9 ma ha sbagliato l’ultima. Cosa dice oggi l’indicatore della Fed?». It has been translated and adapted for an international audience by the Money.it International desk.
