Forecasting And Risk Management Of Renminbi Exchange Rate

Exchange rates are of great importance to international economic association. Asound exchange rate level and a stable exchange rate system are important to economicgrowth, balance of payments and employment of a nation. So the research on exchangerate is one of the important questions of economics.There are a lot of branches in modern western exchange rate theories. The analyses ofexchange rate determination of these theories are based on different assumptionsrespectively. In these theories, the determination of exchange rate is closely related toeconomic fundamental variables, such as relative money supply, interest rate, relativeprices, economic growth, substitutability of foreign and domestic assets and theadjustment speed of equilibrium price. According to the research approaches, thesetheories mainly fall into two categories: the approach based on flows view such as thePurchase Power Parity and Equilibrium of Balance of Payments and the approach basedon assets market such as Flexible Price Model, Sticky Price Model and thePortfolio-balance Model. Some of the theories are opposite to each other, and some ofthem are complementary. But none of them is widely accepted to explain the exchangerate successfully.A large number empirical analyses show that no theory of structural exchange ratedetermination has yet been found that performs well in prediction experiments. Theforecasting ability of them is near to nothing within two years forecast horizon. Only veryseldom has the simple random walk model been significantly outperformed. So a numberof researchers have pursued new approaches to investigate the way of exchange ratedetermination. Empirical research indicated that the relations between exchange rate andfundamentals and exchange rate itself are nonlinear dependent. Several approaches havebeen developed to improve the quality of existing structural exchange rate models byincorporating nonlinear relations into the models. The pessimism about the predictionquality of exchange rate models became generally accepted after the publication of theinfluential paper by Meese and Rogoff(1983). Also Meese and Rose end with a negativeconclusion that incorporating nonlinearities into existing structural models of exchangerate determination does not at present appear to be a research strategy which is likely toimprove dramatically our ability to understand how exchange rate are determined. In the last decade, the neural networks (NN) technology have been increasinglyemployed to study financial and economic problems. Essentially, NN technology is amathematic method. One main advantage of using NN method is that NNs are universalapproximators which can approximate a large class of functions with a high degree ofaccuracy, while most of the commonly used nonlinear models cannot. There are threemajor applications of NNs in economics: first, the prediction of time series, and second,the classification of economic agents, third, modeling bounded rational economic agents.NNs provide valuable tools for building nonlinear models of data with complex structurethat cannot be described precisely by usual methods. Therefore, some researchers employNN technology to predict exchange rate and achieve some success. Recently, the issue of Renminbi exchange rate has become a focus of internationalattention. China's exchange rate policy is facing more and more challenge than before. Ina long period, the pricing of Renminbi to US dollar was based on the balance of domesticand foreign commodity price or the average cost of earning a US dollar in exports, whichin nature are methods of the PPP. The exchange rate of Renminbi was merged with theForeign Exchanges Adjustment Rate in1994. Although Renminbi exchange rate ismarked to US dollar nowadays, the USD/RMB exchange rate may have somerelationship with the price levels of the two countries. During the early days after the collapse of the Briton Woods System, the theory ofpurchasing power parity was proved did not hold to explain the fluctuation of bilateralexchange of main industry countries. With the development of econometrics technologiesand the increasing of samples, it has regained the attention of economics and the attitudetowards the PPP theory changed. In this thesis, the author investigates whether some linear combination of thenon-stationary variables in the three variables equation newly derived form PPP theory isstationary by applying cointegration technology. The result shows that the relationbetween USD/RMB exchange rate and CPIs of China and America is not cointegrationand PPP does not hold on USD/RMB exchange rate. In views of the nonlinearitycharacter of exchange rate, the author establishes a nonlinearity cointegration test basedon NN technology to investigate the nonlinear relation among the above mentioned threevariables. The main idea of this test is train a NN to approximate the nonlinearity relationbetween exchange rate and CPIs of the two countries. Then test whether the residualsseries is stationary by ADF test, calculates the t-statistic. For the high qualityapproximation ability of NN, it trends to overfit the data and approximates the noise well,thus covers the real character of the residuals. To solve this problem, the author select NNfactors for a particular case and calculate the critical values under the null hypothesis ofno cointegration relation through 1000 pairs of independent random walk series. Thisthesis use the same NN structure to regress the relation between USD/RMB exchangerate and CPIs of China and America, find that there exists nonlinearity cointegrationrelationship among the three variables. We can draw the conclusion that the CPIs couldexplain the fluctuation of exchange rate to some extent. Therefore, the author establishesan NN forecasting model which has the CPIs as forecasting variables to forecastexchange rate. The forecast horizons we chosed is 1,6 and 12months as Meese andRogoff(1983). We use a benchmark model of random walk without drift and apply an Stest to distinguish the difference between the two models. The empirical result shows thatthe NN model is better in out-sample forecasting than the RW model and the differencebetween the two models is significant. According to the above idea, this thesis test andforecast HKD/RMB, GBP/RMB and JPY/RMB exchange rate, and find there are nononlinearity cointegration relationships between the three exchange rates and CPIs of thetwo countries and region respectively. Long-term memory of time series is a nonlinear style of a stationary time series, it isuseful to long range time series forecasting. In this thesis, the author investigates whetherlong-term memory present in the return series of USD/RMB, HKD/RMB, GBP/RMB andJPY/RMB exchange rate by employing the modified R/S approach. The thesis finds thatlong-term memory only presents in USD/RMB exchange rate return series. The authorshares the view that there is nonlinear dynamics in exchange rate series, so this thesisonly establishes short term forecasting models using NNs. With the input units vary from1 to 4 and the hidden units vary from 1 to 5, there are 20 different NN models to beapplied to each exchange rate series. The author selects lagged values of return as inputvalues, so the models the thesis adopts are nonlinear AR models. The structure of NNmodels the author adopts is different from commonly used MLP which contains weightsdirectly form input units to output units, and the mapping of the models have two parts oflinearity and nonlinearity. So we may distinguish the nonlinear structure within the timeseries by the weights of nonlinear part. Simply using forecasting errors to evaluate the performance of models will neglectsome useful information. In financial forecasting, economists found that directionalaccuracy and profits appears to be closely related while the correlation between criteriasuch as root mean squared error (RMSE) and actual profits is not significant. So in thisthesis, the author uses RMSE and the success ratio of directional accuracy (SR) as criteriato evaluate the performance of forecasting models. An DA test could be used to seewhether or not SR values the models get differ significantly from SR values that wouldbe obtained in the case that the observed value and the forecast value are independent.Therefore, by taking SR as an criterion we could overcome the shortcoming that theresults of NN models can not be examined by statistic test. In this thesis, the author compares the results of NN forecasting models with linearAR models and RW models. Evaluated by the two criteria, the forecasting performance ofNN models the thesis select is better than both linear AR models and RWs. SR of NNmodels could achieve up to 75% accuracy and significant in statistic test. Therefore thethesis draws the conclusion that NN models may have power in forecasting return ofexchange rate series and the application of NN models is appropriate as the analysis ofthe weights of nonlinear parts of the models shows there may be significant nonlinearstructure in exchange rate series. After 1970's, interest rate and exchange rate fluctuation became increasingaccompanied by the loosening of financial supervision and the development of financialliberalization. With the increasing of financial market factors fluctuation, market risk hasbecome an important part of financial risk. The management of exchange rate risk plays acritical role in management of market risk. Value at risk (VaR) technology is developed inthe situation that the international financial market is facing tremendous volatility. It canmeasure the risk of a financial company in a general framework and become an importanttool and effective approach for financial companies and regulators. Economic researchershave developed various approaches to calculate VaR such as parametric, historicalsimulation, and Monte Carlo approaches. In this thesis, the author uses an neural network method called mixture densitynetworks (MDN) to forecast the VaR of the Renminbi exchange rate risk of a portfoliocontains three foreign exchange rate assets. MDNs are suitable tools for modelingconditional densities, so the author uses historical 5-day and 30-day moving averages andmoving standard deviations to forecast the conditional density functions of the selectedportfolio. The conditional density functions MDN get are mixture normality density...