台灣50指數於民國91年10月29日起正式對外公布並開放交易ETF後至今,因可直接參與臺灣市場整體表現之理想工具,並且進行實際的股價指數交易買賣,其交易量逐漸成長並成為今日重要的投資標地物,進而對預測台灣50指數趨勢便成為重要的議題。總結過去對台灣50指數的預測技術中,以非線性模型的運用較為廣泛,如基因演算法、類神經網路等,或以時間序列模型GARCH等方式。 本研究是針對倒傳遞類神經網路運用在預測台灣50指數報酬率的方法上,提出改進的技術,並針對所模擬的股價報酬率,提出不同的檢測方法。研究中採用50指數成分股與技術指標作為預測的自變數,針對民國95年一月一日至96年十一月三十日,共週資料100筆及日資料495筆作實證。本研究採用命中率的方式來檢驗預測的效果,不同於過去在外部預測的樣本上(Out-Of-Sample)以誤差值為標準的判斷方式。同時加入移動視窗的做法,分析類神經網路的訓練區間長度與時間移動行為所產生的影響。並且期望透過改良的技術達到提升命中率或增進命中率的穩定程度,避免命中率高的現象只是偶然發生的結果。 研究結果顯示,穩定命中率的方式可以從兩方面著手。一為類神經網路模型的選擇、參數的設定、轉移函數的選擇以及運用快速學習的演算法。二為採用變數篩選的技術。最後採用改良後的類神經模式以技術指標為輸入變數證實,其外部預測的效果良好。
After October 29, 2002, TSEC Taiwan 50 index has rose officially and announced outward and traded ETF. Because it is a fine instrument to participate trading Taiwan whole market performance directly, its volume of trade grows up and becomes important today. Predicting Taiwan 50 index trends become an important subject. The techniques for predicting Taiwan 50 index are focus on non-linear model, such as gene algorithm、neural network...etc., or time series model, for example, GARCH. This research provide some methods to reformed backpropagation neural network on forecasting TSEC Taiwan50 Index Returns, and bring a different examine measure for simulate stock index returns. In the research, it introduce compositions of Taiwan 50 index and technique index for independent variables, and adopt 98 weeks, from January 1, 2006 to November 30, 2007, to the research. Hit ration is used for inspection on the outcome of forecast. Not the same as it use MSE as standard criterion on out-of-sample forecasting inspection in the pass. This study adds the method of Moving Window to analyze the effect of the length in training period and time shift backward at the same time. It expect to use reformed backpropagation neural network to improve or stabilize hit rations, which prevent the high hit ration taking place is just accidental. In this research, it find two ways to stabilize hit rations. First is chose the parameters、transfer functions and use variable learning rate algorithm. Second is use sieving variables out before neural network. Finally, we test the reformed neural network used by technical indices, and it show that stable.