The performance comparisons of various time series and artificial intelligence approaches in the return predictions of Taiwan stock market

• Main Authors: Jui-hsin Lau, 劉瑞鑫
• Other Authors: Tsung-Nan Chou
• Format: Others
• Language: zh-TW
• Published: 2003
• Online Access: http://ndltd.ncl.edu.tw/handle/26755509795817791487

碩士 === 朝陽科技大學 === 財務金融系碩士班 === 91 === This paper employs time-series econometric model and other artificial intelligence approaches to predict Taiwan Stock market index return. The approaches applied in this study include the back-propagation neural network (BPN), genetic algorithm based neural network (GANN1), retrained genetic algorithm based neural network (GANN2) and back-propagation neural network with GARCH parameter (GARCH-BPN). The purpose of GANN1 model is to avoid the over-fitting of the BPN model and reduce the effort of training process convergence. The GANN2 model modifies the weight and bias values of the GANN1 model through retraining and is more suitable for the characteristic of data set. The GARCH-BPN model involves the impact of the conditional variance of returns within the neural network to predict returns. The performance measurement criteria of this study consist of the forecast error, directional accuracy and forecast encompassing test to compare the estimating and predicting ability of different approaches. The results show that using neural network to retrain the weight and bias values which evolved by genetic algorithms is better than other time-series and artificial intelligence models in most performance measurement criteria. The ARMA model displays moderate performance about forecast error, but inferior to other artificial intelligence models, and reveals the time-series model in virtue of statistical assumptions and restrictions which differ from artificial intelligence model finding the corresponding relationship betweens inputs and output values. During forecast encompassing test, the GANN2 encompasses others models, denotes the GANN2 model can capture more information of the data set than other models.