Credit analysis using a combination of fuzzy robust PCA and a classification algorithm

Abstract

Classification is a key part of credit analysis and bankruptcy prediction and new powerful classification methods coming from artificial intelligence are often applied. Most often classification methods require pre-processing of data. This paper presents a two-part classification process that combines a pre-processing step that uses fuzzy robust principal component analysis (FRPCA) and a classification step. Combinations of three FRPCA algorithms and two different classifiers, similarity classifier and fuzzy k-nearest neighbor classifier, are tested to find the combination that gives the most accurate mean classification result. Tests are run with a small Australian credit data set that can be considered “rough” and to require “robust” methods, due to the small number of observations. The created principal components are used as inputs in the classification methods. Results obtained indicate a mean classification accuracies of over 80% for all combinations. It becomes clear that parameters of the used methods clearly affect the results and emphasis is put on finding suitable parameters.