A Deep Learning Hybrid Model for Enhanced Credit Score Prediction

Abstract

Determining someone's creditworthiness accurately is still challenging, especially if they have a short credit history or mostly use cash. In these situations, traditional credit scoring techniques frequently fall short, which could cause misclassification and financial losses for lenders. To improve credit score prediction, the research focuses on creating a hybrid deep learning model that blends Recurrent Neural Networks (RNNs) and Deep Neural Networks (DNNs). This study aims to evaluate the role of behavioral and traditional data in the existing credit scoring models, develop a hybrid deep learning model that integrates both data types for predicting credit scores, validate the developed model, and create a web-based tool to visualize the model. Data preparation techniques include feature engineering and feature selection to find complex patterns in the data. Design Science Research (DSR) is the research design used for developing artifacts in this study. The hybrid RNN+DNN model outperforms solo RNN and DNN models, as shown by performance evaluation measures like accuracy, precision, recall, F1-score, AUC-ROC, confusion matrices, sensitivity, specificity, MSE, and RMSE. With an AUC-ROC score of 0.7971, it attains balanced and dependable credit score predictions, with the lowest RMSE (0.723) and MSE (0.523) and sensitivity of 0.8372 for class 2 and specificity of 0.8790 for class 0. By offering consumers a useful interface, the web-based tool designed to display the hybrid credit scoring model increases the usefulness of the research. This application makes it easier to input financial data, analyze it so that the hybrid RNN+DNN model can use it, and then display the anticipated credit scores ('Good,' 'Standard,' or 'Poor'). Streamlit's features guarantee a smooth user experience, confirming the proposed model's efficacy in practical situations. However, difficulties with interpretability, computing requirements, dataset quality, and ethical issues are mentioned. More extensive and more varied datasets should be obtained, hyperparameters should be optimized, computational efficiency should be increased, interpretability should be strengthened, and the model should be validated against actual credit scoring systems in the real world. By addressing these issues, hybrid deep learning models will be further improved, guaranteeing their ethical use in credit evaluation as well as their scalability, comprehensibility, and reliability. This will also help marginalized communities become more financially included.