People want to know if a recession is coming, and the usual answers are useless. One economist says yes, another says no, a headline says a downturn is imminent, and a week later another says the coast is clear. None of it gives you anything to act on. A recession probability model exists to replace that noise with a single, honest number, the odds that a recession begins within a given window of time. This is how that number gets built, how it earns trust, and what it can and cannot tell you.
The core idea: from signals to a probability
A recession probability model does one thing. It takes a set of economic signals, measured today, and turns them into the probability that a recession starts within some future period, three months out, a year out, two years out. Instead of a yes or no, you get a percentage, an 18 percent chance within twelve months, a 45 percent chance within twenty four.
The reason to express it as a probability is honesty. Nobody can know the future with certainty, and any model that claims to is lying. Recessions are genuinely hard to predict, and even the best methods get some wrong. A probability is the truthful way to state that. It says, given everything the signals are showing right now, here is how likely a downturn is, acknowledging that likely is not the same as certain. A number tells you more than a verbal guess precisely because it’s specific and because it can be checked against what actually happens.
Why leading signals, and not the obvious numbers
The first real decision in building one of these models is which signals to feed it, and the answer runs against intuition. The numbers most people watch, jobs, GDP, consumer spending, are the wrong inputs for prediction, because they’re coincident or lagging. They tell you where the economy is now or where it just was. By the time unemployment is clearly rising and GDP is clearly falling, a recession is already underway, which is too late to be a warning.
So a prediction model is built instead on leading indicators, the signals that move before the economy turns. Decades of economic research, from the Federal Reserve banks to the IMF and the ECB, keep landing on the same short list of signals that carry real predictive power. The slope of the yield curve, the gap between long and short term interest rates, is the most studied and most reliable, with a track record of inverting before recessions going back more than half a century. Credit spreads, the extra yield lenders demand on risky debt, widen when financial stress builds. Business investment, measured through new orders for capital goods, falls when companies lose confidence in the months ahead. And the broad stock market, which prices expected future earnings, turns down when investors collectively see a slowdown coming. These are forward-looking by nature. They reflect decisions and expectations about the future, which is exactly what you need to forecast one.
Why combine signals instead of trusting one
You could build a model on the yield curve alone, and the New York Fed famously publishes one that does. But the research is consistent that combining several leading indicators outperforms any single one. The reason is that each signal has blind spots and false alarms of its own. The yield curve can be distorted by unusual central bank policy. The stock market panics at things that never become recessions. Credit spreads can spike on a scare that passes. Any one of them, read alone, will cry wolf.
Combining them cancels out a lot of that noise. When one signal flashes a warning for a reason unrelated to the broader economy, the others stay calm, and the combined reading reflects that disagreement rather than overreacting. When several independent signals turn negative at once, that agreement is far more meaningful than any single alarm, because it’s much less likely that four unrelated measures are all being fooled by the same fluke at the same time. A composite is more stable, more accurate, and less jumpy than its parts. That’s a repeated finding across the literature, that indexes combining several measures beat individual indicators at signaling recessions.
How the number actually gets calculated
Under the hood, models like this typically use a statistical method built for yes-or-no outcomes. The most common is called a probit, or the closely related logit, regression. The idea is simpler than the name. The outcome being predicted is binary, either a recession begins in the window or it doesn’t, a one or a zero. The model looks across decades of history and learns the relationship between the signal values at a given moment and whether a recession followed within the window. It finds, in effect, how much each signal should move the odds, and in which direction.
Once that relationship is estimated from history, you feed in today’s signal values and the model returns a probability, its best estimate of how likely a recession is given conditions that look like these, based on how often recessions followed similar conditions in the past. It’s not magic and it’s not a crystal ball. It’s a disciplined way of asking, when the signals looked like they do now, how often did a recession actually come.
Why it reads across multiple time horizons
A single recession probability, a 12-month number, is useful but incomplete. The stronger approach reads several horizons at once, three, six, twelve, eighteen, twenty four months, because the shape across those horizons carries information that no single number can.
Different signals lead by different amounts. The yield curve tends to warn a long way out, often a year or more before a recession. Other signals move closer to the event. So reading only the twelve month figure throws away what the near and far horizons are saying. The full term structure, the whole set of probabilities across time, tells you not just whether risk is elevated but where in the cycle the economy sits. A reading that is low in the near term and high further out describes risk gathering on the horizon while the present stays calm, the classic shape before a downturn. A reading elevated across every horizon describes a cycle close to turning. A reading calm everywhere describes clear skies. The shape is the message, and a model that reports only one horizon hides most of it.
How you know it works: testing out of sample
A model that fits history perfectly is worthless if it only works in hindsight. The real test is out-of-sample performance, which means checking whether the model would have worked using only the information available at the time, without peeking at the answer.
The honest way to do this is to walk through history as if living it. At each past month, show the model only the data that existed then, ask it for a probability, and then check what actually happened next. If the model consistently raised its recession odds before recessions it was never shown, it has genuine predictive power. If it only looks good when it can see the whole timeline at once, it’s just describing the past. Serious models are validated this way across many decades and multiple recessions, and the ones worth trusting are the ones that gave real warnings ahead of downturns in that blind, out-of-sample test, not just a clean fit after the fact.
What a model like this cannot do
Honesty about the limits is what separates a credible model from a doom machine. A recession probability model cannot predict shocks that originate outside the economy. A pandemic, a sudden war, a financial accident with no economic buildup, none of these leave a trail in the leading signals beforehand, because they don’t come from the business cycle. The 2020 recession is the clearest example, a model reading the cycle had no way to see a virus coming.
It also produces false alarms. Elevated odds are not a guarantee, they’re elevated odds, and sometimes the recession that the signals warn about doesn’t arrive, because the Fed changes course or conditions improve. A model that reads 45 percent is saying a downturn is a real possibility, not a certainty, and being honest about that means accepting it will sometimes be high when nothing happens. What a model reads is the building business cycle, the slow accumulation of stress in rates, credit, investment, and expectations that precedes a normal recession. That’s a lot, and it’s more warning than most people ever get, but it isn’t everything, and a model that pretends otherwise shouldn’t be trusted.
Why a probability beats a prediction
The value of all this is what it gives you that a yes-or-no call never can, time and proportion. A probability that climbs across months lets you prepare in steps rather than react in a panic. The cheap moves, building a little cash, shortening commitments, holding off on an expansion, are available early, while the odds are still rising and conditions still look fine, and they vanish once a downturn is obvious to everyone. A model that reads the signals early turns that window from something visible only in hindsight into something you can act on with time to spare. It won’t tell you the future. It will tell you the odds, honestly, early enough to matter.
See where the signals stand
The recessionodds.com model updates monthly with the probability of a U.S. recession from 3 to 24 months out, read from five leading public signals, tested out of sample across more than fifty years and eight recessions. Subscribe to get the reading in your inbox when it moves.
recessionodds.com is published for informational purposes only and is not financial, investment, or legal advice.