Modelling and Forecasting Stock Returns:Does Neural Network Model Perform Better than GARCH Model?

在股價報酬率預測上過去常使用統計方法研究資料，但統計方法對於資料有諸多限制，例如假設資料為線性、殘差項為白噪音等等。股價報酬資料含有高狹峰、拒絕常態分配、波動叢聚之特性，因此過去文獻常使用GARCH模型來預測股價報酬率資料，而近期研究則表示類神經網路模型點預測結果更為精準，故本研究將比較兩種模型之預測優劣。另外，若僅採用點估計並無法描繪資料之變異性，故本研究也同時將焦點著重於，採用GARCH模型與類神經網路模型之預測區間的建構及描繪，用以補足點估計之缺漏。 本研究使用S&P500、Nasdaq 、DJIA三個股價資料，來檢驗GARCH模型與類神經網路模型，兩者在股價報酬率上之點預測及區間預測的預測結果，並且進一步比較兩組不同解釋變數之預測優劣。實證結果顯示，在點預測上GARCH模型有較低的預測誤差，在區間預測上GARCH與類神經網路模型因採用的解釋變數不同，造成最佳預測模型有所不同。GARCH模型採用半徑、中位數為解釋變數，可得到高資訊量及覆蓋率的預測區間；若採用最高報酬率、最低報酬率為解釋變數時則建議採用類神經路模型來建構預測區間。Statistical methods have often been used in literature to predict stock price return, but the past statistical methods have their own limitations such as it might assume the data is a linear relation and residual is a white noise. As stock price return contains characters of leptokurtic, reject normality assumption, and volatility clustering; therefore in past literature, GARCH model had been used to predict stock price return datasets. However, in recent literature has shown the neural network model is more accurate. In addition, the primary problem for point prediction is it cannot properly handle the variability with datasets. As a result, this study aims at describing the interval prediction of GARCH model and neural network model in order to compensate the disadvantages of point prediction. The objective of this study is to compare GARCH model and neural network model for stock price return rate of point prediction and intervals prediction respectively by using S&P500, NASDAQ and DJIA stock price; and make further efforts on discussing the prediction performance between two sets of explanatory variables. According to obtained results, GARCH models have smaller predicting error on point prediction; in the case of interval prediction, it has different outcome by adopting different explanatory variables of GARCH models and neural network separately. Furthermore, GARCH model can generate well-informative and coverage probability by adopting Radii and midpoint as a dependent variable; while using the highest return rate and the lowest return rate, neural network model greatly outperforms than GARCH model.第一章 緒論 1 第一節 研究背景與動機 1 第二節 研究目的 3 第二章 文獻回顧 5 第三章 研究方法 11 第一節 GARCH模型 11 第二節 類神經網路模型 12 第三節 性能指標 14 一、點預測性能指標 15 二、區間預測性能指標 16 第四章 實證分析 19 第一節 資料來源及處理 19 第二節 模型選擇 20 一、線性模型的選擇 20 二、類神經網路模型的設定 20 第三節 實證結果 21 第五章 結論 23 附錄 25 參考文獻 45 一、中文部分 45 二、英文部分 4