Recession Probability Models: How Economists Forecast Downturns

Recession probability models convert current economic data into numerical probability estimates of recession over a specified time horizon (typically 12 months). The models combine one or more leading indicators with statistical or econometric techniques to produce probability outputs that financial-media coverage and policy discussions widely cite. The most widely tracked models include the Federal Reserve Bank of New York's yield-curve probit model, Bloomberg's monthly economist consensus survey, internal models from major investment banks (Goldman Sachs, JPMorgan, Bank of America), and the Atlanta Fed's GDPNow current-quarter GDP nowcast. As of the October 2026 release (based on data through September 2026), the publicly computable model, the NY Fed yield-curve probit, reads 12.76 percent for 12-month-ahead recession probability.

Put 12.8 percent in context using the model's own published history rather than a remembered rule of thumb. Across the full series since 1960 the mean reading is 15.3 percent and the median is 9.1 percent. So the current reading sits just under the average month and still above the median one. It has fallen a long way from the cycle peak of 70.9 percent in May 2024, the highest since 1982, reached as the twelve-month lag worked the deep 2023 inversion through the model. Proprietary bank models and the Bloomberg economist survey also publish 12-month estimates, but those are subjective or paywalled and are not independently verified here. The verifiable picture is recession risk a little under its long-run average, neither elevated nor unusually low.

The NY Fed Yield-Curve Probit Model

The Federal Reserve Bank of New York's recession-probability model is based on the 10-year minus 3-month Treasury spread, originally developed by economists Arturo Estrella and Frederic Mishkin in research published through the 1990s. The model is published monthly at the New York Fed website and is widely cited in financial-media coverage of recession risk. The methodology is publicly documented and reproducible, distinguishing it from many proprietary investment-bank models.

The technical specification uses a probit regression to convert the monthly average yield spread into a probability estimate of recession within the next 12 months. The estimation is based on monthly data since 1959. The yield-spread coefficient is the dominant input: in the months since 1985 when the spread sat between +1.4 and +1.6 percentage points, the probability twelve months out came in between 6.1 and 7.8 percent, averaging 6.9 percent. Probabilities rise as the spread narrows or inverts. The deepest monthly inversion of the last cycle was negative 171 basis points in May 2023, and the reading twelve months later, in May 2024, was 70.9 percent, the highest in the series since the early 1980s.

The current reading of 12.76 percent reflects the yield-curve un-inversion. The 10Y-3M spread averaged +0.96 points in September 2026 on the New York Fed's measure and stood at +0.88 on 19 October (FRED T10Y3M), well above the inversion threshold. The probability has been falling since the May 2024 cycle peak as the re-steepened curve worked through the model's twelve-month lag, and now sits under the series mean of 15.3 percent, above its 9.1 percent median.

What the Model's Track Record Actually Is

The model's reputation runs ahead of its record, and the record is checkable: the New York Fed publishes the whole monthly series back to 1960 in allmonth.xls. Reading that file against the NBER chronology, for each of the eight recessions that began after 1968, here is the highest reading the model reached in the twelve months up to and including the business-cycle peak month.

Our reading of the New York Fed's published monthly series (allmonth.xls, October 2026 release) against the NBER business-cycle chronology. Checked 20 October 2026.

So the model cleared 30 percent ahead of five of the eight, and in two of those five only in the peak month or the month before. It never approached 30 percent before the 1973-75, 2001 or 2020 recessions. Where it did lead, it led usefully: a full twelve months before the July 1981 peak, four months before December 2007. The honest summary is that this is a real signal with an uneven record, not a dependable twelve-month alarm, and it has no useful history at all against supply-shock or pandemic recessions.

The other side of the ledger is false signals. The model ran above 30 percent continuously from late 2023 into the first half of 2026 (its 2026 high was 30.5 percent in April) without a recession arriving. That is the largest false-signal episode in its history, and it is the main reason to treat the current low reading as information rather than reassurance.

The Bloomberg Economist Consensus

Bloomberg surveys approximately 75 professional economists each month, asking each economist for a 12-month-ahead recession probability estimate. The published consensus is the median of the responses (with mean and quartile data also reported). The economists range from major Wall Street investment banks (Morgan Stanley, Goldman Sachs, JPMorgan, Bank of America, Citigroup) to academic economists at universities, to economists at the IMF, World Bank, and OECD.

The survey provides a useful aggregation of professional judgement that incorporates information beyond what any single quantitative model captures. Economists draw on multiple inputs (yield curve, jobless claims, manufacturing PMI, consumer confidence, geopolitical conditions, fiscal-policy stance, credit markets) and their estimates reflect both quantitative analysis and qualitative judgement about how the various factors interact. Because the Bloomberg consensus sits behind a paywall and cannot be independently verified here, this page does not quote a specific current figure for it; historically it has tended to run somewhat higher than the yield-curve model, reflecting the mixed signals across underlying indicators.

The historical performance of the Bloomberg consensus is reasonable but not exceptional. Economist surveys tend to be conservative, avoiding extreme readings (above 70 percent or below 10 percent) even when conditions might warrant them. The consensus also tends to be backward-looking, adjusting estimates after the underlying economic conditions have changed rather than anticipating changes ahead of time. The 2007-09 recession was preceded by Bloomberg consensus readings around 30 percent (relatively elevated for the survey) but the consensus did not exceed 50 percent until the recession was clearly underway.

Investment Bank Internal Models

Major investment banks maintain proprietary recession-probability models that combine multiple inputs in firm-specific frameworks. Goldman Sachs Research publishes its model output through periodic notes to clients (with a roughly monthly publication cadence); JPMorgan Asset Management and Bank of America publish similar outputs. These readings sit behind client paywalls and are not independently verifiable here, so this page does not quote specific current percentages for them. Historically they have tended to cluster somewhat above the publicly computable yield-curve model rather than below it.

The proprietary character of investment-bank models limits external evaluation of methodology. The general framework typically involves multiple inputs (yield curve, jobless claims, ISM manufacturing PMI, consumer confidence, credit spreads, financial conditions) combined through regression or factor-model techniques. The exact weights and time-horizon specifications vary by firm. The models are typically updated monthly or quarterly with revised input data.

The dispersion across investment-bank models tends to be modest (typically within 5 to 10 percentage points), reflecting that the underlying economic conditions are reasonably similar to the inputs that the public Fed and Bloomberg models use. Substantial divergence between investment-bank models would suggest either methodological differences mattering or proprietary information being incorporated; in practice, both occur to a limited degree but the headline outputs typically cluster.

The Atlanta Fed GDPNow Nowcast

The Federal Reserve Bank of Atlanta publishes the GDPNow nowcast, a current-quarter GDP estimate updated continuously as new data becomes available throughout the quarter. The nowcast is not a recession-probability model in the traditional sense (it does not produce a probability over a future horizon) but rather a real-time estimate of the GDP figure that will eventually be reported for the current quarter.

The GDPNow methodology uses a Bayesian state-space framework that updates the GDP estimate as new monthly indicators are released (PCE deflator, retail sales, construction spending, ISM manufacturing data, factory orders, employment situation). The estimate evolves through the quarter from an initial broad estimate to a more precise estimate as more data accumulates. The final GDPNow estimate (typically published a few days before the official BEA advance estimate) has a strong historical track record of approximately matching the BEA initial GDP estimate.

The latest GDPNow reading for the third quarter of 2026 is +3.7 percent annualised (FRED series GDPNOW, 30 September 2026), so the quarter is tracking well clear of contraction. For context, BEA has Q1 2026 real GDP growth at +2.5 percent and Q2 at +2.2 percent. (For the live figure, see the Atlanta Fed page linked below; it is revised continuously through the quarter.) Recession-monitoring application of GDPNow involves watching for sharp downward revisions during a quarter (which would suggest a developing contraction) and for two consecutive quarters of negative GDPNow estimates (which would partially satisfy the two-quarter rule even before official BEA data is published). Neither pattern is currently present.

The Sahm Rule as Real-Time Coincident Indicator

The Sahm rule is a coincident indicator rather than a probability model. It triggers a binary signal (recession yes or no) when the three-month moving average of the US unemployment rate rises 0.50 percentage points or more above the lowest three-month average from the prior twelve months. The current reading is 0.00 (September 2026), far below the trigger threshold, having receded from its July 2024 crossing of 0.50 (no recession followed) as unemployment came down from 4.5 percent to 4.1-4.2 percent.

Some practitioners interpret the Sahm rule reading as roughly equivalent to a probability estimate (a 0.40 reading might be interpreted as a 40 percent recession-imminent reading). Sahm herself has cautioned against this interpretation, noting that the rule is designed to function as a binary trigger rather than a continuous probability. The cleanest framing is that the Sahm rule and the probability models are complementary: the probability models forecast the chance of recession over a forward window, while the Sahm rule and other coincident indicators (jobless claims, manufacturing employment, real personal income) provide real-time monitoring of whether the recession that the probability models are forecasting has actually begun.

The combination of the yield-curve model (12.76 percent) and the Sahm rule reading (0.00, far below trigger) provides a consistent picture: no recession-imminent signal from either. The yield-curve reading is an ordinary one rather than a reassuring one, since the model has averaged 15.3 percent since 1960 with a median of 9.1 percent. If unemployment resumed rising and the Sahm rule climbed back toward 0.50, the probability-model readings would likely rise as well. As long as the labour market holds and the curve stays positive, the readings stay subdued.

Why the Models Disagree

Different models use different inputs, different time horizons, and different statistical methodologies, all of which can produce different numerical estimates from the same underlying economic conditions. The NY Fed yield-curve model uses a single input (the 10Y-3M spread) with a long historical track record but is purely backward-looking econometric estimation. Bloomberg consensus aggregates judgement that incorporates many inputs but is less mechanically reproducible. Internal investment-bank models combine multiple inputs in proprietary frameworks. The Atlanta Fed GDPNow nowcast tracks current-quarter GDP estimates rather than predicting future recession.

The dispersion across models is itself useful information. When models cluster tightly, the consensus signal is stronger and the confidence in the underlying assessment is higher. When models diverge, the underlying economic conditions are more ambiguous and individual model readings should be interpreted with greater caution. With the publicly computable yield-curve model now at 12.8 percent, the verifiable signal points to recession risk near its long-run norm rather than elevated. The dispersion was much wider in 2023-24, when the yield-curve model published readings of 60-70 percent while the labour-market indicators showed no recession, reflecting the central interpretive question of how much weight to assign to the unusually deep yield-curve inversion.

The 2024-26 False-Signal Episode

This has been the hardest stretch in the model's history. Its readings ran above 30 percent in every month from November 2023 to November 2025, peaking at 70.9 percent in May 2024 and still as high as 61.8 percent in August 2025, and no recession arrived in any of those predicted windows. By early 2026 the series had come back to the 20s, and the latest reading is 12.76 percent for September 2027. Bloomberg consensus and investment-bank models published their own elevated readings over the same period, but those sit behind paywalls and are not quoted here.

The false-signal episode is the largest in the modern history of these models. The yield-curve-based models were particularly affected because the 2022-25 yield-curve inversion was the longest in the modern monitoring period. The Bloomberg consensus and investment-bank models were less acutely affected because they incorporate broader inputs that did not all align with the yield-curve signal. The Atlanta Fed GDPNow has not produced false signals because it tracks current-quarter conditions rather than forecasting future recession.

The implications for future model interpretation are significant. If the 2024-26 episode is a one-off (driven by the unusual COVID-era cyclical dynamics that have not occurred before), the models may continue to function reliably going forward. If the episode reflects structural changes in the modern services-dominant US economy (manufacturing weight in the LEI, service-sector resilience to monetary tightening, household-balance-sheet buffers from COVID-era stimulus), the models may need methodological refinements to function as effectively in future cycles. Either interpretation will be testable as more data accumulates through 2026 and beyond.

Practical Application of Probability Readings

Probability readings are inputs to decisions, not decisions themselves. A 30 percent recession-probability reading is meaningful but does not by itself determine appropriate household, business, or policy responses. The relevant question is what action thresholds different stakeholders should apply. For households, modestly elevated recession probability might motivate maintaining or modestly increasing emergency-fund balances, accelerating high-interest debt repayment, and reviewing employment portability and skill-development priorities. For businesses, elevated readings might motivate maintaining cash reserves, postponing major capacity expansion, and ensuring credit-line availability. For investors, elevated readings might motivate rebalancing toward defensive asset classes for those near retirement, but not for younger investors with long-horizon dollar-cost-averaging strategies.

With the yield-curve model at 12.8 percent, recession probability is around its baseline historical norm, the range where routine financial prudence is warranted rather than defensive positioning. If the readings rose back to 60-70 percent in coming months, more substantial adjustments would be appropriate; for now the verifiable signal does not call for them.

For the underlying single-indicator deep dives, see the Sahm rule, the yield curve indicator, initial jobless claims, ISM manufacturing PMI, and the Conference Board LEI. For the live current 2026 dashboard, see current probability 2026.