Robust logistic regression for SMEs' default prediction

Abstract

Predicting SMEs' default and financing promising firms means protecting 99\% of all enterprises in the EU, as well as the largest part of the European value added and jobs. Accordingly, there is a vast literature studying SMEs' default in European Countries, mainly based of accounting indicators. Logistic regression is the benchmark model for classification of default, due to remarkable performances comparable with those of machine learning methods, with an immediate interpretation. The goal of the thesis is to search for alternative methods such as robust logistic regression to predict SMEs' default in Italy. Firstly a comprehensive bibliographic research on SMEs' default prediction is carried out, followed by the description of the collection and creation of a large dataset of balance sheets downloaded from Aida database. Thereafter the available libraries in R are used to apply robust logistic regression to classify defaulted firms within the collected data, rearranging the functions where needed. Lastly, a comparison of classification rates, the significance and relevance of the coefficients with the standard logistic regression outcome is performed to contextualize the results within the relevant literature. The aim is to point up that although new methods should be taken into consideration, the logit model remains the cornerstone of credit risk evaluation, besides credit scoring.