Combining Wavelet Decomposition and Wavelet Neural Network for Non-Stationary Financial Time Series Forecasting
|關鍵字:||時間序列;小波轉換;神經網路;股市預測;非定性;time series;wavelet transform;neural network;stock market forecasting;non-stationary|
Traditional time series analysis methodologies are based on probability and statistics with the assumption of stationary and linear properties. However, the system dynamic of time series usually arise with highly nonlinear and non-stationary properties, these conventional time series forecasting models cannot satisfy the feasibility and accuracy of which research desires. Consequently, the “multi-resolution wavelet neural network hybrid forecasting model,” capable of adaptive forecasting non-stationary time series, is proposed in this research. The original time series is decomposed into subsequences in different resolution scale using the wavelet decomposition, which is efficient in processing chaotic signals. Furthermore, combined with the wavelet neural network architecture, which is referred to an universal function approximator, to establish the time series forecasting model, and expect this model to predict accurately and conquer the restriction of the traditional models when encounter non-stationary time series. The TAIEX of Taiwan stock market index is used for one and five day ahead forecasting of close price and price change to demonstrate the proposed model. Two other forecast results, one is obtained from traditional autoregressive model, and the other is without using the wavelet decomposition, were also used to compare with the proposed model. The experimental results indicate that the multi-resolution wavelet neural network hybrid forecasting model can accurately predict non-stationary financial time series and provide a valuable reference for making investing decision.
|Appears in Collections:||Thesis|