An Unsupervised Learning Approach Towards Credit Risk Modelling Using DFT Features and Gaussian Mixture Models

Abstract

One of the most important problems in the present times is to estimate the risk in lending financial resources with respect to its returns. Credit risk can harm a lender by increasing collection costs and causing cash flow inconsistency. Lenders use credit risk modelling to assess the amount of credit risk associated with lending credit to borrowers. Financial statement analysis, default likelihood and machine learning are the options for credit risk analysis models. And, solving this kind of problem using machine learning techniques is known as credit risk modelling. In this process, we fit data having loads of features related to the financial conditions of a person into the model to classify a lender as a defaulter or non-defaulter. In this study, we used an unsupervised machine learning technique for this task. First, we applied two feature selection methods, viz. using the Pearson’s correlation coefficient and chi-square test, to select certain features which are less informative for the task. Feature selection is one of the pre-processing standards in designing advanced solutions because it does not only alleviate dataset dimensions but also improves a model’s performance measures. We also applied the fast Fourier transform (FFT) algorithm to get the discrete Fourier transform (DFT) of all the selected features, as supplementary and artificial feature vectors to the model. To deal with class imbalance, we used a specific variant of synthetic minority oversampling technique (SMOTE), i.e. Gaussian-SMOTE. Finally, we applied Gaussian mixture model (GMM) at distinct values for its parameters like the ‘covariance type’. With this and the optimally selected parameters, our methodology was able to achieve a classification accuracy of 85.48%.