VAR Method Based On GARCH Model Measured Foreign Exchange Risk

Since China has reformed RMB exchange rate system on July21,2005, RMB exchange rate changes increasingly frequently. After our country joined WTO, the share in international currency and capital markets become larger. Now, China is the country with the largest foreign exchange reserves, but under pressure from international political forces and economic factors, RMB is faced with the pressure of appreciation, as a result, the foreign exchange reserves will decline rapidly in international foreign exchange market. In such circumstances, the management of foreign exchange risk is particularly important.VAR is the mainstreaming in risk management methods, which is used by many financial institutions and companies, especially Western developed countries. In China, the method is used mainly in bond market and rarely used in the foreign exchange market. Therefore, the article tried to find a suitable method by using VAR method which is based on various model of GARCH according to the statistical characteristics of RMB exchange.Based on the statistical characteristics we found that the data has the characteristic of thick tail, non-normality, asymmetry, dependence, volatility cluster, time-varying conditional variance sand long memory, etc. The models of GARCH can eliminate the heteroscedastic effect. In order to forecast the maximum return and the maximum loss, the VAR value based on various of GARCH models in high position95%,99%and low position5%,1%,0.5%and0.25%were calculated, and the accuracy test was made on the value of VAR:With the increase of the confidence level, the overestimation of the underside VAR value (underestimate risk) by all kinds of GARCH models based on normal distribution became more and more serious, which illustrated that the left tail of normal distribution was too thin; in the high position95%and99%, various of GARCH models based on normal distribution, T distribution and GED distribution overvalued VAR values, while various of GARCH model based on skewed T distribution were basically passes the test, which further illustrated that the sequence of R has the properties of the left partial. That is consistent with the trend that dollar compared to RMB is devaluation; Although upper VAR value estimated by the kinds of GARCH models based on the skewed T distribution were closest to the expected values, with increasing confidence levels, underside VAR values estimated by the kinds of GARCH models based on the skewed T distribution were more and more smaller than the expected values (overestimate risk), which illustrated that the left tail of skewed T distribution was much thicker than the R sequence. According to the predicted results, kinds of GARCH models based on T distribution and GED distribution could forecast the value of VAR stablely. With increasing the confidence level, kinds of GARCH models based on T distribution overestimated VAR values (underestimated risk) within a certain degree, indicating that the left tail of T distribution was a little thinner than R sequence; All kinds of GARCH models based on GED distribution underestimated the VAR value (overestimate risk), and with the increase of the confidence level, the degree of overestimation was more and more bigger,which indicated that the left tail of GED distribution was thicker than R sequence. In the risk management process, risk is more important than profit, so the model which predicted the underside VAR most accurately was the optimal model. Generally speaking, the model whose mean equation had conditional standard deviation was better. Consider the close degree between the forecasted underside VAR values and the expected values, the GARCH-M model based on T distribution is best and with the increase of the confidence level, the predicted values were relatively stable.In order to compare with GARCH models, VAR value of low position based on ARMA model were calculated, and the accuracy test were made. VAR values calculated based on GARCH-M model of T distribution are far better than the VAR values calculated based on ARMA model.