Prediction of Credit Defaults based on Weight Dimensionality Reduction Neural Network and M-Band Wavelet Transform

When companies decide to give out loan money to a person or another company, it is of utmost importance that the money they gave out will be returned to them. Lenders need a way to predict whether or not that person or company will default and not give their money back. In this day and age, there is a lot of data regarding companies and people and way too many of them to perform a careful analysis to determine if they will default or not. The most optimal solution for the task at hand will be a pre trained system that receives financial information and returns a credit default prediction regarding the other party.. We aim to solve the problem of Credit Defaults with Machine Learning. Firstly, we implemented M-band discrete wavelet transform (MDWT) to decompose our dataset into M different frequency components to discover some hidden information that we would not get otherwise. In this paper we propose a novel weight dimensionality reduction neural network (WDR-NN) that uses dimensionality reduction techniques such as UMAP, Wavelets and PCA to generate a new set of neural network weights and then pass relevant information that is in a reduced dimension but preserves the overall structure of the networks weights. In our research, two Datasets were used: England Companies Binary Classification of whether the company went bankrupt at some point and Moody's and Fitch Credit Defaults using binary classification to see if their rating by the agency was higher than a given threshold. The results have shown that our WDR-NN model outperformed a standard neural network by yielding a 13% accuracy increase in predicting Company fraud as binary classification. We utilize Shapley Values on our WDR-NN and find that Operating Cash Flow Share and Days of Sales outstanding are the two most important features in determining a company's default. We conclude by demonstrating that our WDR-NN outperforms traditional methods such as Least Discriminant Analysis, Logistic Regression, Decision Trees, and Support Vector Machines in various metrics on both datasets.