Stock Market Forecasting Based on Wavelet Neural Network

• Main Authors: Yi-pu Chen, 陳苡溥
• Other Authors: Kuei-fang Chang
• Format: Others
• Language: zh-TW
• Published: 2011
• Online Access: http://ndltd.ncl.edu.tw/handle/58217882740498805356

碩士 === 逢甲大學 === 應用數學所 === 99 === 　　Traditional time series analysis must establish the data into a stationary and linear situation to obtain a better forecasting result, but the financial data is often in a non-stationary and non-linearly status which restricts the result. This thesis applies the WNN on stocks, funds and raw material prices to forecast two weeks, and then compare the RMSE of the two week forecast value by three aspects: different forecasting techniques, different data records and different wavelet function. By the experiment, we learn that the RMSE in two weeks are controlled by the four variables of percentage of training data, the number of training data, the initial of network parameters and the threshold. 　　First, HHT and DWT employ AR(2) forecast and to compare the forecasting value of RMSE in two weeks, the input document utilized in this stage is the stock prices of 50 representing companies in Taiwan. Results show that 46 out of 50 stocks in WNN forecasting value surpasses the former, only 4 groups of Taiwan Cooperative Bank, President Chain Store Co., Far Eastern New Century Corporation, and Quanta Computer Inc., have inferior WNN forecasting value to the former. Secondary, using WNN to the stocks, funds and raw materials price 100 to 500 input data carries on RMSE of two week forecast value. The results show that wavelet neural network do not need to limit the amount of the input data and trends to project fine results from all data, significantly promoting the convenience in the usage. Better forecasting results come out when percentage of training data is less than or equal to 20%. Finally, under the same network construction and input document''s condition, calculates 2 wavelet functions psi1 and psi2 in 4 groups based on 100 to 500 input data two-week forecast value, and compares its RMSE, respectively. When WNN takes for the wavelet function psi1 and psi2 , the results show that about 3 groups of minimum value of the RMSE of psi1 is smaller than that of the RMSE of psi2 .