Feature selection and fuzzy rule-based classifier applied to bankruptcy prediction in banks

Abstract

We propose a fuzzy 'if-then' rule-based expert system preceded by feature selection for predicting bankruptcy in Turkish, Spanish and the UK banks. The system comprises three phases: feature selection, rule generation and optimisation. Feature selection reduces the number of rule antecedents, thereby enhancing human comprehensibility without sacrificing accuracy. Top five features are selected using wavelet neural network-based algorithm, t-statistic and f-statistic. Then, the reduced set of features is fed to a fuzzy rule-based classifier (FRBC) that generates rules and the third phase optimises the rule base. Tenfold cross-validation is performed throughout. The frequency of the presence of a rule in each fold decides the common rule base. The experiments are conducted for 2-4 partitions and area under receiver operating characteristic curve criterion selects the optimal partition and the best feature selection algorithm. Results indicate the effectiveness of FRBC with feature selection in predicting bank bankruptcy.