This paper critically examines the conventional practice in financial economics of using the 10-year--three-month Treasury yield spread to predict recessions without empirical verification of this specific maturity pair selection. The authors employ a machine learning approach to systematically identify potentially superior maturity pairs and coefficient specifications that might enhance recession forecasting accuracy.

The comprehensive analysis reveals that while machine learning algorithms can achieve likelihood gains in model fitting, these improvements do not translate into statistically significant enhancements in out-of-sample predictive performance. The primary limitation identified is estimation error, which prevents the machine learning approach from delivering meaningful forecasting advantages over the traditional spread measure. The findings demonstrate robustness across various forecasting horizons, control variables, sample periods, and data treatments including oversampling techniques for handling imbalanced recession observations. Ultimately, the research provides empirical support for the continued use of the conventional 10-year--three-month Treasury yield spread in recession probability forecasting, suggesting that the established measure remains statistically optimal despite the availability of more sophisticated computational methods.