PREDICTION EFFICIENCY WITH ARTIFICIAL NEURAL NETWORK FOR SET50 INDEX

Abstract

This study aim to forecast the time series of an Artificial Neural Network (ANN) of the SET50 Index. The time series data is comprised of the closing prices of the SET50 index and trading volume and technical factors from 1 January 2012 to 31 December 2016. One thousand one hundred and eighty -nine daily samples are employed. The data for the last 244 days is used to calculate the forecasting performance of the ANNs.

In the study, multi-layer perception (MLP) models are used. Hidden layers and result layers Back propagation is used to determine and design an artificial neural network. Using the results of the experiments and the training of this system to generate the SET50 Index price forecast from the 39,304 network model trainings, the network model [13,15], [14,12], [ 13,7], [16,11], [15,4], giving the root mean square error (RMSE) of 6.7666, 6.9298, 7.5773, 7.3192, 7.4431. The five network models for forecasting 244 days out of the sample. The next day was repeated for 244 days, then the MAPE values ​​of the five networks were compared to find the lowest MAPE network.

The results of the established models consisted of 5 structures of the artificial neural networks: [13,7] , [15,4], [14,12], [16,11], and [16,11] is efficient model for forecasting SET50 Index with Mean Absolute Percentage Error (MAPE) values of 1.7943, 2.3449, 3.1136, 3.4516 and 3.6244. The [13,7] network was the most accurate network in the 244-days’ forecast. It was found that the artificial neural network model can be highly predictive because it is a non-parametric model. (Non-Parametric) using the learning principle of the model to establish relationships of weight and structure within the network. This is very effective and effective in establishing data relationships. Models with more hidden layers. It does not imply that the predictive ability is greater than the model with a sufficiently adequate hidden layer. It is clear that the two-hidden layer network is enough to solve the problem. Adding more layers Does not affect better forecasting performance. However, the number of neurons in each stratum has an effect on the predictive efficiency. By the way, the good network structure model. There should be Neural Network in the last layer less than the previously hidden layer.