本研究嘗試以倒傳遞類神經網路，進行台股加權股價指數期貨隔日收盤指數預測，研究期間由2000年1月至2010年12月，採用基本面指標與技術指標做為輸入變數，加入不同區隔樣本期間資料方法計有景氣循環法、時間移動法、空間移動法。
1.在輸入變數參數設定值方面，威廉指標5日參數比10日、BIAS 5日比9日預測效果更佳。
2.在研究期間內，如利用區隔樣本資料期間，景氣循環法、時間移動法、空間移動法，皆可提高預測績效結果。
3.在研究期間內，如採用景氣循環指標做為類神經網路輸入變數，不論是否利用逐步迴歸法減少輸入變數，皆無法提高預測績效結果。
4.利用空間移動法選取的訓練期間，樣本內數列較平穏，預測績效結果較佳。
由以上研究結果可知，倒傳遞類神經網路模式也可捕捉期貨不規則波動，有不錯之預測績效。
This study uses back-propagation neural network (BPNN) theory to forecast the next day closing index of TAIEX from January 2000 to December 2010. We use fundamental indicators and technical indicators as input variables. The learning sample is divided into business cycles, sliding windows, space windows.
The result shows that:
1. For the input parameters, using the 5-day William’s Indicator accompanies with the 5-day bias predicts better than using other parameters or using the indicators separately.
2. The predicting result can be more accurate, if we divide the period of sample data into business cycles, sliding windows, and space windows during the training phase.
3. Using business cycle as the input parameter for BPNN cannot improve the predicting performance, no matter whether using the stepwise regression method to reduce the number of input parameters or not.
4. The forecasting result is better when using longer training space window as the sample data is more stable.
