Forcasting Performance of Artificial Neural Network and Time Series : Empirical Evidence from the Price of Polaris Taiwan Top 50 Tracker Fund

• Main Authors: Juei-Ming Hsiao, 蕭瑞銘
• Other Authors: Po-Chin Wu
• Format: Others
• Language: zh-TW
• Published: 2011
• Online Access: http://ndltd.ncl.edu.tw/handle/b93s79

碩士 === 中原大學 === 國際貿易研究所 === 99 === This study is the use of Taiwan 50 ETF pricing as a predictive analysis. Three hundred and thirty-four trading records could be made a sample data which be picked up from “Taiwan Economic Journal Co., Ltd” database and date period is from Jan.2nd, 2004 to June 25th, 2010. Use Eviews 6.0 and NeuroSolutions 5.17 to be analysis tools. By using the method of Moving Window, All data has been divided into six parts and each part has been moving four weeks. Data in each part is subdivided into “ training data “and “ testing data “. According to “ training data “ in each part, we can set up appropriate Time Series and Hybrid Neural Network analysis models to evaluate forecast performance by analyzing data that pick from outside samples. Weekly closing price is the only one variable used in the evaluating process to simplify the Time series model. By time moving, observe the forecast quality and stable of AR and GARCH of traditional Time Series and Hybrid Neural Network model. This paper is not the same as others to use many variables to increase forecast ability. In order to examine the model forecast performance, forecast error rate, correct movement rate and rate of return are used for evaluated standards. Several findings can be confirmed as follows： In forecast error rate standard, Time series model is better than Hybrid Neural Network model. If we check six parts each by each, we can find the forecast result is unstable when the vibration of the data that are picked up out of models is too large. In forecast correct movement rate and rate of return standards, Hybrid Neural Network model is better than Time series model and rate of return exceeds actual stock price return. If we check six parts each by each, we can find the forecast performance is lower and hard to have conclusions when the vibration of the data that is picked up out of models is too large. This research can explain why the forecast results are not the same of sample data and data is picked from outside sample in other articles. The reason is the unstable data that is picked from outside sample. When we do data forecast In the feature, we have to use the method of Moving Window and then we can get trusted conclusions.