The Research On Combinative Forecasting Of Exchange Rate Based On Wavelet Analysis And Artificial Neural Network

Exchange rate is the rate at which one currency can be converted into another. It plays a decisive role in the flow of foreign capital, international banking businesses, foreign capital reserves, international trade and risk management. Since the exchange rate regime reform on July 21, 2005, the government has softened exchange rate controls step by step, which consequently led to unprecedented volatility of the RMB exchange rate. Therefore, it is of great importance to gain a deeper understanding of the behavior of exchange rates to reveal its innate running mechanism and ultimately improve forecasting capabilities.This paper firstly reviews the development of exchange rate forecasting models and techniques, and specifically analyzes the application of neural networks in exchange rate forecasting. There is a brief review on the techniques for time series denoising and the usage of wavelet analysis in denoising fields, especially in exchange rate forecasting field. With consideration of the non-linear and noise-contaminated characteristic of the exchange rate series, together with the advantages of neural network and wavelet analysis in non-linear forecasting and denoising respectively, a combinative model based on wavelet analysis and artificial neural networks for exchange rate forecasting (WDANN) is developed. This model consists with two parts: Decomposing and re-constructing the exchange rate by wavelet, to separate noise from information; The other part is training the neural network with the denoised series, then using the trained network for forecasting exchange rates.The WDANN model is tested by empirical research with the RMB/USD exchange rate. The experiment results shows that the out-of-sample performance is improved on the ANN model with denoising process. The denoising performance of wavelets with different wavelet functions, different decomposing levels and different wavelet thresholds is also examined by experiments. The results indicate that the following parameters yield more accurate forecasts: coifN wavelet function, decomposing level of three, and sqtwolog threshold.