How recessionodds.com works
A forward read of U.S. recession risk, built from five public economic signals and tested honestly against more than fifty years of history.
What the model is
Recessionodds.com reports a single thing: the probability that a U.S. recession begins within the next 3, 6, 12, 18, and 24 months. That set of five numbers is a calendar of risk, low and near or high and far, that you read as a shape rather than a single figure.
It is driven by five public signals, each free and published by the government or the markets: the yield curve, the recent direction of that curve, credit spreads, new orders for business equipment, and the stock market. Every one is sourced and linked on the home page, so nothing is hidden and anyone can check the inputs.
How it is tested
A forecasting model is only worth anything if you measure it on data it has never seen. This one is validated with a walk-forward, out-of-sample method: starting decades ago, the model is fit only on the data available up to each month, then scored on the next month it has not yet seen, stepping forward one month at a time through history. The accuracy reported is the accumulation of those genuine, unseen predictions, never a model graded on the same data it learned from.
By that standard the model separates pre-recession periods from normal ones with an out-of-sample accuracy in the range of 0.92 to 0.94 AUC, which is strong by the standards of the recession-forecasting literature, and its strength holds across the whole calendar rather than fading at the long horizons where most signals go quiet.
What it saw before each recession
Here is the honest record, the real out-of-sample readings the model would have produced in the year before each downturn in its validation window. The failures are included, because a track record that hides them is marketing, not evidence.
The 1980 recession
The model was already loud a year ahead, climbing to roughly 99 percent by late 1979 as the curve inverted hard going into the Volcker era.
The 1981 to 1982 recession
Its clearest call. The signal reached nearly 100 percent by mid-1981, well before the deepest downturn of the postwar period took hold.
The 1990 to 1991 recession
A harder turn to see, and the record shows it: the warning rose to about 82 percent in late 1989, a real signal but a more modest one than the early-1980s calls.
The 2001 recession (dot-com)
Caught early and clearly, reaching about 95 percent by the end of 2000 as the late-1990s boom rolled over.
The 2008 financial crisis
The signal climbed to roughly 84 percent by March 2007, about nine months before the recession officially began. 2008 was a credit-driven collapse that the yield curve alone times poorly, and the strength of this warning came from the credit, orders, and stock-market channels catching the buildup the curve was missing. It is the clearest case for using more than one signal.
The 2020 pandemic recession
This one is included for honesty, not as a win. No model built on economic data can foresee a virus. The reading was elevated going into 2020, but that was the residue of a 2019 yield-curve inversion that had nothing to do with the pandemic, not a true detection of the shock that came. The model reads the building business cycle, not a bolt from the blue.
Where it fails, plainly
The model produces false alarms. The loudest in its history ran from 2022 to 2024, when the longest yield-curve inversion in decades pushed the long-horizon reading very high and no recession followed on the historical schedule. Tested directly, false alarms cannot be cleanly separated from real warnings in advance, because at a distance the approach to a recession and a false alarm look alike. That is not a flaw to apologize for; it is why the sensible response is to prepare cheaply and early and commit heavily only when the near term confirms it.
Two more honest limits. The model is built on only about eight recessions, a small number of real events, so every accuracy figure carries a wider band of uncertainty than a clean number suggests, and no single reading should be treated as a prophecy. And it cannot anticipate shocks from outside the economy, as 2020 showed.