Research Of G-VaR Prediction Based On RMB Exchange Rate Risk

Since the "811 Exchange Rate Reform",the volatility range of RMB exchange rate has gradually increased,and the volatility risk of exchange rate has increased the difficulty of political decision-making,so it is of great significance to manage and control exchange rate risk.In 1993,J.P.Morgan proposed VaR,a risk measurement under the framework of linear expectation theory,but VaR lacks the ability to measure losses in extreme cases.The emergence of nonlinear expectation theory guides the improvement direction of risk measurement in the classical probability space.The nonlinear expectation theory can better describe the dynamic change in the real world than the classical linear expectation theory.Therefore,this thesis mainly researches the risk measurement under the framework of nonlinear expectation theory——G-VaR.At present,G-VaR is an efficient index to measure stock risk,but the exchange rate as an important macroeconomic variable is greatly influenced by the economic policy uncertainty,G-VaR prediction method can’t analysis economic policy uncertainty on the volatility of exchange rate from multiple time scale,thereby it limits the G-VaR forecast accuracy.Therefore,this thesis proposes to integrate EEMD method into the framework of G-VaR prediction method.It is assumed that EEMD method decomposes the exchange rate return series into independent components under sublinear expectation.On this basis,the influence of economic policy uncertainty on each component is deeply analyzed.This thesis takes USD/CNY,EUR/CNY and JPY/CNY as examples,the EEMD method is used to achieve the independent decomposition of exchange rate return series,assume that exchange rate return series which is decomposed into three independent components under sublinear expectation——the high frequency,low frequency and residual.For the stationary high frequency component,ARMA-GARCH model is used to extract its mean and volatility,the economic policy uncertainty index is taken as exogenous variable,and GARCH-MIDAS model is used to predict its mean and volatility.The mean and volatility of low frequency component are extracted using the simple moving average model,and the forward rolling prediction is made using the Attention-LSTM model.For the non-stationary residual series,the simple moving average model is also used to extract its mean and volatility,and the ARMA model is used to predict them.Furthermore,according to the predicted value of volatility,the optimal window is selected to calculate the predicted value of upper volatility and lower volatility.The independent properties of the high frequency,low frequency and residual guarantee the linear additivity of mean,upper volatility and lower volatility under sublinear expectation,so the linear sum of the predicted value of mean,upper volatility and lower volatility of each component can get the corresponding parameter predicted value of the original sequence.Finally,the G-VaR can be obtained by the G-VaR prediction formula.In order to highlight the advantages of the G-VaR prediction method based on EEMD proposed in this thesis,this thesis conducts comparative experiments with other G-VaR prediction methods,and carries out empirical analysis of insample prediction and out-of-sample prediction respectively.Binomial test,Kupiec test,Christoffersen independence test and Bernoulli trial likelihood ratio test are selected as evaluation indexes to evaluate the prediction performance of the method.Finally,the method proposed in this thesis shows good robustness and universality in both in-sample and out-of-sample prediction.