Abstract
Credit Prediction requires analysis of multiple user-level features including their previous transactions, spending patterns, billing patterns, previous line of credit status, etc. To analyse the risk associated with credit sanction, various Machine Learning Models (MLMs) are proposed by researchers, but very few of them are capable of confidently estimating true credit limits for different users. Thus, banking firms & financial advisory firms are not able to confidently recommend credit amounts on a per user basis. To overcome these limitations, this paper discusses design of a novel Credit prediction model that can perform Risk Minimization via Ensemble Learning process. The proposed model uses a large-scale dataset that combines user’s personal details like Education Level, Gender, Marital Status, and Age with details about previous spending & payment patterns to estimate credit default probability, along with maximum amount that can be guaranteed with minimum risk & maximum retention rates. The model uses combination of Naïve Bayes (NB), Support Vector Machine (SVM), Deep Random Forest (DRF), Multilayer Perceptron based Neural Network (MLP), & Logistic Regression (LR) to predict user’s credit scores. These scores are combined with a correlation-based model which assists in estimation of maximum amount of credit that can be extended to the customer with confidence of higher retention rates. Due to a combination of these methods, the proposed CPRAMLCT Model is capable of high accuracy, high precision, better recall, and low delay operations. The classification models were combined using a unique union-based ‘mode’ method, which assisted in identification of common & correctly classified samples. This reduces classification errors, which assists in improving classification accuracy by 6.5%, recommendation precision by 15.5%, recommendation recall by 8.5%, and recommendation speed by 14.5% under different real-time scenarios. Due to these performance enhancements, the model is useful for a wide variety of applications.