Abstract

A precise and reliable loan status prediction is of the essence for financial institutions, However, the lack of real-world data and biases within that data can greatly impact the accuracy of machine learning models. Another challenge faced by loan status prediction models is class imbalance, where one category (such as approved loans) is much more common than another (such as defaulted loans), leading to skewed predictions towards the majority class. This study inspects Generative Adversarial Networks (GANs) to augment the data and improve the machine learning models’ performance. Several machine learning (ML) models including but not limited to Support Vector Machines (SVM) and ensemble bagged trees were employed on a Kaggle loan dataset (380 samples). Baseline training and testing accuracies were 86.9% and 86.3% (SVM) and 84.5% and 82.1% (ensemble). ActGAN (Activating Generative Networks) was then utilized to generate synthetic data points for both accepted and rejected loans. Retraining the models with new augmented data showed remarkable improvements: SVM accuracies for training and testing rose to 94.4% and 93.4%, while ensemble models achieved 97.4% and 95.8%, respectively. Other ML models were also explored such as KNN, Decision tree and logistic Regression and showed promising results in terms of accuracy as compared to the state of art. These findings put forward that GAN-based data augmentation can enhance the performance of loan status prediction. Future research could explore GAN’s impact of different architectures and assess the general applicability of this approach.