The inversion of the yield-curve has long held up as the single most prominent early-warning indicator of a looming recession in the United States. 

Yet, recessions are arguably the result of a complex convolution of many economic variables. While such a setting strains the capabilities of orthodox models, machine-learning algorithms are supposed to excel in such environments. Still, classical probit and penalized regression models turn out to be resilient competitors. 

Beyond providing a plain point-forecast and quantifying its uncertainty, I address the criticism of ML’s limited interpretability. The results corroborate the standing of the yield-curve as the principal predictor of U.S. recessions, followed by labor- and stock-market indicators. 

Bundling predictive ability and interpretability within a single package, I propose RecÆ, a structural autoencoder-type architecture that leverages the predictive power of the yield-curve while conditioning its relation with the probability of an upcoming recession on the state of the economy. The RecÆ identifies three consecutive regimes with differing sensitivities of recession-probabilities to a steepening of the yield-curve. The results advocate for carefully designed ML to be a valuable addition to the econometric toolbox.