Securities such as stocks and bonds are examples of most peoples’ investment instruments. If there was a way to predict tomorrow’s closing price, one can draft the investment strategy today in advance, and prepare for essential funding. Therefore, the objective of this study is to predict the next day's closing price.
This study uses back-propagation neural network model to predict and compare the performance data with "multiple regression model". The research target is "Polaris Taiwan Top 50 Tracker Fund", a.k.a. the "Taiwan 50", which is an Exchange Traded Fund (ETF). We use 10 years historical data as the training dataset, and use one month out of sample period to verify whether it is correctly predicted.
In the previous literature, very few researches simultaneously split "training information" and "test period data", if using neural networks to predict the stock index. In this study, we divide the training data into seven parts, and split the test data into five datasets, producing a total of 35 combinations of forecasting models. We then apply the "time axis shift method" on seven models using "moving window method" to shift 100 consecutive trading days in order to collect the prediction results. Finally, we analyze the generated data to obtain the best forecasting model.
The results confirmed that prediction effectiveness of the "back-propagation neural network model" is better than the "multiple regression models". Our best model can predict the rise or fall of the next day closing price with 56% accuracy using the “moving window method”.
