Predicting bankruptcy of local government: A machine learning approach

In this paper we analyze the predictability of the bankruptcy of 7795 Italianmunicipalities in the period 2009-2016. The prediction task is extremely harddue to the small number of bankruptcy cases, on which learning is possible.Besides historical financial data for each municipality, we use alternative insti-tutional data along with the socio-demographic and economic context. The pre-dictability is analyzed through the performance of the statistical and machinelearning models with a receiver operating characteristic curve and the precision-recall curve. Our results suggest that it is possible to make out-of-sample pre-dictions with a high true positive rate and low false-positive rate. The modelshows that some non-financial features (e.g. geographical area) are more im-portant than many financial features to predict the default of municipalities.Among financial indicators, the important features are mainly connected to theDeficit and the Debt of Municipalities. Among the socio-demographic charac-teristics of administrators, the gender and the age of members in council areamong the top 10 features in terms of importance for predicting municipaldefaults.