An AutoML application to forecasting bank failures
Άννα Αγραπετίδου, Paulos Charonyktakis, Periklis Gogas, Théophilos Papadimitriou, Ioannis Tsamardinos
Abstract
We investigate the performance of an automated machine learning (AutoML) methodology in forecasting bank failures, called Just Add Data (JAD). We include all failed U.S. banks for 2007–2013 and twice as many healthy ones. An automated feature selection procedure in JAD identifies the most significant forecasters and a bootstrapping methodology provides conservative estimates of performance generalization and confidence intervals. The best performing model yields an AUC 0.985. The current work provides evidence that JAD, and AutoML tools in general, could increase the productivity of financial data analysts, shield against methodological statistical errors, and provide models at par with state-of-the-art manual analysis.