Assess deep learning models for Egyptian exchange prediction using nonlinear artificial neural networks

Abstract

Financial analysis of the stock market using the historical data is the exigent demand in business and academia. This work explores the efficiency of three deep learning (Dl) techniques, namely Bayesian regularization (BE), Levenberg–Marquardt (lM), and scaled conjugate gradient (SCG), for training nonlinear autoregressive artificial neural networks (NARX) for predicting specifically the closing price of the Egyptian Stock Exchange indices (EGX-30, EGX-30-Capped, EGX-50-EWI, EGX-70, EGX-100, and NIlE). An empirical comparison is established among the experimented prediction models considering all techniques for the time horizon of 1 day, 3 days, 5 days, 7 days, 5 days and 30 days in advance, applying on all the datasets used in this study. For performance evaluation, statistical measures such as mean squared error (MSE) and correlation R are used. From the simulation result, it can be clearly suggested that BR outperforms other models for short-term prediction especially for 3 days ahead. On the other hand, lM generates better prediction accuracy than BR- and SCG-based models for long-term prediction, especially for 7-day prediction.