This paper presents a new dataset on the dynamics of non-performing loans (NPLs) during 88 banking crises since 1990. The data show similarities across crises during NPL build-ups but less so during NPL resolutions. We find a close relationship between NPL problems—elevated and unresolved NPLs—and the severity of post-crisis recessions. A machine learning approach identifies a set of pre-crisis predictors of NPL problems related to weak macroeconomic, institutional, corporate, and banking sector conditions. Our findings suggest that reducing pre-crisis vulnerabilities and promptly addressing NPL problems during a crisis are important for post-crisis output recovery.