Rmb Exchange Rate Based On Neural Network Prediction

Recently, RMB/USD exchange rate was a topic to people because of that USA wanted to classify China as exchange rate controlling country. So to research exchange rate is meaningful. In this paper, the method of nonlinear time series and neural network are used to analysis the series of RMB/USD exchange rate.In the empirical studies part, a sample that is central parity of RMBA/USD from 1994.3.14 to 2010.4.14 is choosed as study for nonlinear test. Series of exchange rate is changed into sequence of fluctuant exchange rate through the first difference to be smooth, which is examined by a spectrum of nonlinear test, such as JB test, autocorrelation test, BDS test and ARCH test. The tests prove that the sequence of fluctuant exchange rate is non-normal, two-stage autocorrelation, non-independent and identically distributed and it has ARCH effect. Since sequence of fluctuant exchange rate is nonlinear, it fits nonlinear model as neural network. A 798 sample from 2007.1.4 to 2010.4.14 is choosed to build a neural network model, the first 768 sample is used for building a net and sample forecast, and the last is forecasting sample. In the process of parameter design in network, the optimal lag is obtained that has three values:6,7,10, which is also the number of input and hidden neurons, based on AC criterion. Then the best training sample is calculated as 768 that is the optimal sample in different freedom according to MCPT method. Based on these parameters, MLP and RNN2 network model are built to fit and predict the sequence of fluctuant exchange rate. MLP(10) is the best model in the first 768 sample, while RNN2(10) is the perfect one to the last 30 sample. Since MLP network may has such problem as "overfitted", RNN2(10) is seen to be the optimal network model to the sequence of exchange rate.In this paper, only two network models are used to forecast exchange rate, and the network model is not compared with ARCH model, which is the weakness of this study.