Initialization of smooth adaptive neural network weights with a cultural algorithm for SET index prediction

Abstract

In the business sector, predicting the movement of the Stock Exchange of Thailand (SET) index is challenging. Due to worldwide stock market fluctuations, investors commonly invest in price-changing businesses solely in the long term. Therefore, an accurate SET index movement prediction method is significant for investment purposes and has been the goal of many previous studies. Some studies have indicated that neural network (NN) models perform more effectively and accurately than traditional statistical models; accordingly, NNs employing backpropagation (BP) with sigmoid and smooth adaptive activation functions (SAAFs) and 10 metaheuristic algorithms to determine the initial prediction weights were developed in this study. An experiment was conducted using a Thailand SET50 index dataset, and the results revealed that the model utilizing SAAFs with a cultural algorithm (CA) for weight initialization yielded more precise and efficient predictions than those of other competing models. This finding indicated the possibility of applying the proposed method for SET index movement prediction in the future.