Enhancing credit score classification: a time-effective deep learning approach

Abstract

Credit Risk Management is a strategic foundation for decision making and long term financial stability in today's banking environment. The purpose of this study was to compare and contrast multiple Multilayer Perceptron (MLP) architectures in terms of their predictive ability as well as their potential for reducing computational latency. By using data sets that ranged in size from 20,000 to 100,000 records, we measured the predictive abilities of these models using Accuracy, F1-Score, Recall and Precision metrics; as well as measuring the amount of time it took to execute each model. The findings of our study showed that although most of the various MLP models we tested were capable of producing similar levels of predictive accuracy, the Larger Batch Size (LBS-MLP) model had the greatest effect in terms of significantly reducing the amount of time it takes to execute. Specifically, at a 100,000 record data set level, the LBS-MLP model was able to complete the processing of the data set in 42.216 seconds, which was significantly less than the baseline model and HA-MLP model, both of which are highly computationally intensive. Therefore, we conclude that the LBS-MLP model provides an optimal frontiers solution for financial institutions that need to produce high precision credit assessments and also have the real-time operational capacity to support the digital lending process.