Comparing Stochastic Gradient Descent and Mini-batch Gradient Descent Algorithms in Loan Risk Assessment

Abstract

Despite the increase in the practice and the advent of artificial neural networks (ANNs) and deep learning (DL) to reduce the number of slippage loans in the banking and financial industries, some of the Nigerian banks and financial institutions are not aware of its efficacy. Loan request assessment would expand credit pronouncement efficiency, and inadvertently save the time and cost associated with loan analysis. This research considered, compared, and contrasted two artificial neural networks algorithms –the Mini-batch (Normal) gradient descent and the stochastic gradient descent algorithms. Each of the algorithms was separately used with back-propagation neural networks to develop loan evaluation models. Samples were collected from a Nigerian bank, these samples contain an index of default and non-default customers, and were used to train the loan evaluation models. The outcomes of the research show that the stochastic gradient descent algorithm outperforms the mini-batch gradient descent algorithm in terms of the percentage accuracy (0.863 and 0.835 respectively) and the space complexity, while the mini-batch gradient descent algorithm performed better in time complexity (107 µs and 1 ms respectively). The stochastic gradient descent algorithm was superior in identifying borrowers that were default with 87% accuracy.