股票市場複雜敏感,且容易受到許多因素影響,由於因素間具有非線性關係且存在許多雜訊(Noise),直接利用原始資料進行預測往往無法得到良好績效,所以可以利用特徵萃取方法(Feature extraction methods),將股價指數資料轉換到特徵空間(Feature space)藉以降低上述特性的影響,提升預測結果的有效性。本研究應用非線性獨立成份分析(Nonlinear independent component analysis, NLICA)技術於時間序列之特徵萃取,以獲得較多的隱含資訊,並以粒子群最佳化(Particle swarm optimization, PSO)做為支援向量迴歸(Support vector regression, SVR)的參數調整工具,建構整合非線性ICA、PSO、SVR之時間序列整合預測模式。本研究以財務時間序列資料為實驗樣本,將建議之NLICA-PSO-SVR整合預測模式與結合其他多種特徵萃取方法之預測模式做預測績效比較,由於NLICA能夠有效的從觀測的資料中萃取出代表股價主要趨勢特徵的獨立成份,因此所提之結合NLICA的整合預測模式之預測績效優於使用LICA、PCA、KPCA等方法做為前處理之預測模式,也優於沒有使用NLICA做前處理的預測模式。此外,結合PSO的整合預測模式能夠有效降低運算與時間成本,在預測結果上也得到了一些提升。
Stock market is complicated and sensitive, which could be easily influenced by many factors. Due to there are nonlinear relationships between these factors and exist noise, using raw data to make predictions often cannot obtain good performance. So, we employ feature extraction methods to transform stock price data into feature space to reduce the drawbacks which above-mentioned and to rise the effectiveness of prediction outputs. In this research, nonlinear independent component analysis (NLICA) is applied as feature extraction method to time series data in order to discover hidden information, and particle swarm optimization (PSO) is applied as parameter-optimizing tool of support vector regression (SVR) to build up NLICA-PSO-SVR integrated time series forecasting model. The proposed model is compared with models which are integrated with other feature extraction methods. Due to NLICA can extract independent components (ICs) which are representing the main trend of stock price from observed data effectively, the performance of proposed model is better than other integrated models. On the other hand, integrating PSO to forecasting model can reduce computing cost and time-wasting dramatically and is also beneficial to forecasting performance.
為了持續優化網站功能與使用者體驗,本網站將Cookies分析技術用於網站營運、分析和個人化服務之目的。
若您繼續瀏覽本網站,即表示您同意本網站使用Cookies。