Publication

Abstract : Predictive machine learning algorithms offer an efficient way to perform mundane computations and analysis that otherwise would have taken lots of time and manual effort. In banking and finance, creditors analyze the risk level of loan products in order to set interest rates for potential clients. In this paper, we examine supervised neural network approaches that can be used to automate the risk assessment process on loan applications. Dataset includes customers with potential features who are classified as risky and non-risky. To avoid loss by granting loans to risky customers, classification is done using multiple adaptative neuron network (Madaline) and multiple-layer perceptron (MLP) models, and loan is granted only to non-risky customers. Predictions of Madaline networks are optimized using parameter tuning methods. MLP model handles label imbalance and skewness of data toward one label. Models are trained tested with two datasets containing records of individual customers and companies as debtors. Prediction performance of Madaline and MLP models is assessed using different measures. MLP model showed higher prediction accuracy of 97% and 84% for customer and company dataset, respectively.