Credit modelling using hybrid machine learning technique

Credit evaluation models are the important tools used by banks for the evaluation of loan customers as good or bad. These models are developed as a part of data mining projects using mainly the Classification and Clustering tasks. Their accuracy plays a very significant role as they are the backbones behind the important decisions taken by banks. The accuracy can be improved by using many factors, some of these are the use of good machine learning techniques, balanced input data, and using hybrid techniques in model development. The machine learning and statistical techniques can be combined in various ways for creating the effective hybrid models. In this paper the input data has been balanced to avoid biased model training towards the larger class. Machine learning techniques which have been proved successful in many experiments on financial data are used for this study. The machine learning techniques used are: Naïve Bayes, MLP, RBF, Logistic Regression and C4.5. First single models have been developed using these machine learning techniques and the one with highest accuracy has been found. Then this model was hybridized with others for improving the classification accuracy. The accuracy of all these models was tested on a separate test set that has not been shown to the model while training. A bench marked credit dataset has been utilized for conducting the experiments. The results of the single and hybrid models shows that the MLP outperformed all other individual models while the hybrid model developed by combining the MLP with MLP gave the best results.