Forecasting the Stock Exchange of Thailand Index Employing Deep Learning by Back-Propagation Neural Network Process

Abstract

with the back propagation is a method in deep learning which is used to estimate the non-linear relationship between input and output. The stock price indexes have a tendency of fluctuating over time, therefore, using a back- propagation neural network method is an interesting way to predict the SET index.
This study aims to forecast the SET index employing back-propagation neural network process and investigate the artificial neural network models that are effective in forecasting the SET Index. The input data by expending closing price data of SET, Dow-Jones, Nikkei, Straits-Times, Hang-Seng Index and daily trading volume of SET since 1 January 2009 to 31 December 2018 total amount 2,437 days. Then the data is divided into 2 groups, the first 2,192 days is for the training data set in order to find the appropriate model for predicting the SET Index, which will be selected from the top 5 models with the lowest mean squared error (MSE). And 245 days remaining, uses for testing the 5 selected models to find the forecasting accuracy with the absolute average percentage error (MAPE). For modeling, it uses a multi-layer neural network which consists of 1 Input layer, 2 Hidden layer s and 1 Output layer, and trained with back propagation.
As a result of this study, in the training part : the models with the lowest 5 MSE are model [8,9,11,1], [8,5,13,1], [8,13,9,1], [8,17,17,1] and [8,12,8,1]. In the testing part, it was found that the above models gives forecasting accuracy with the MAPE of 0.690149%, 0.895387%, 0.763672%, 0.788011%, 0.728610% and 0.895387% respectively. So the model [8,9,11,1] has the lowest MAPE of 0.690149%. Prediction of the SET Index during that period with this model has very close results. Therefore, this model is quite effective for forecasting the Stock SET Index