線性迴歸分析與類神經網路應用於配適與預測財務性時間序列

Abstract

以機械式的交易方法應用於財務性時間序列上已被廣泛的討論，我 們以線性迴歸和類神經網路兩種方法分別對台灣股票市場的21種股票作配 適與預測，並比較其報酬率。在線性迴歸方法上引用了統計學的假設檢定 以制定一套簡易的交易法則，並對各股的價、量、值作報酬率的比較。在 類神經網路方法上我們引用小波轉換以濾除各股價格上的雜訊，以利類神 經網路在追隨股價趨勢上的精確性。從模擬實証中發現線性迴歸方法較有 一定的報酬率，而類神經網路方法在獲利上則較低，顯見其在預測能力上 的劣勢。最後我們提出一項評估新交易方法的簡易法則，並對目前的方法 提出未來改善的方向。 It has been extensively discussed on mechanical trading methods applied in financial time series. In this thesis, we try to fit and forecast the twenty-one stocks from the Taiwan stock market by linear regression and neural networks approaches, respectively, and then compare their profits. In linear regression approach we use the statistical hypothesis test to form a simple trading rule, and apply this rule to the price, value and volume of each stock to compare their profits. In neural network approach we filter out the noise on the price by wavelet transform to approximate the price trend more accurate. In the simulation results, the linear regression approach has higher profit than the neural network approach due to its poor forecastability. Finally, we provide a simple criterion to survey a new trading method and suggestions to improve the two approaches in the future.