Deep Convolutional Neural Networks Versus Multilayer Perceptron for Financial Prediction

Abstract

This paper presents a new approach to apply and evaluate Deep Learning (DL) Convolutional Neural Networks (CNN) versus Multilayer Perceptron (MLP) for financial prediction. We have designed and evaluated a credit scoring model based on neural network classifiers in two variants: (a) MLP with eight layers; (b) DCNN with thirteen layers (six main layers and seven secondary layers). The experiments have used the German credit dataset and the Australian credit dataset. The model performances are evaluated by the following indices: Overall Accuracy (OA); False Alarm Rate (FAR); Missed Alarm Rate (MAR). The experimental results have confirmed the effectiveness of the proposed approach, pointing out the significant advantage of DCNN over the MLP. For German credit dataset, the DCNN leads to the best OA of 90.85%, versus the corresponding best MLP performance of only 81.20%. For Australian credit dataset, the DCNN has led to the best OA of 99.74%, while the MLP has obtained the best corresponding OA of 90.75%.