Weighted Smoothing Support Vector Machine And Empirical Study Of Exchange Rate

As an important part of international finance, exchange rate is always the hot topic for discussion and research. Since July 7 of 2005, China has implemented a new exchange rate system that adjusts the exchange rate on basis of the supply and demand of the market as well as a reference of currencies. It is a supervisory floating system. The new system leads to an increase of frequency and unstable fluctuation of exchange rate, which has negative influences on our financial system. These lead to an increase of the exchange rate risk and the instability of financial market, and pricking up speculation. Therefore, it is indispensable to establish a prompt and accurate exchange rate forecast model to capture the fluctuation of RMB/USD exchange rate in certain future period.Firstly, this paper discusses the development of exchange rate forecasting methods, the application of Support Vector Machine (SVM) in forecast and its limitations. Then this paper describes the theory of SVM, Smoothing SVM Algorithm and introduces the theory of Locally Weighted Regression (LWR). Next, by analyzing the limitation that SVM has the same penalty constant (C) for different training points in the actual application, we do some improvement in the model. Meanwhile, the original optimization problem with constrains is transformed to an unconstrained convex quadratic programming problems. Therefore, a new solution is obtained, which is called Weighted-Smoothing Support Vector Machine (w-SSVR). And its efficiencies are proved by two simulations. On basis of the analyzed classic theories of foreign exchange rate and the characters of RMB, in the fourth part of the thesis, 10 macro economic indexes of China and USA are presented as the factors of influencing RMB/USD exchange rate. Then a W-SSVR exchange rate forecast model is established based on these 10 structured variables. To apply the model, data from Jan. 1999 to Mar. 2008 are quoted to make empirical research on exchange rate. The result indicates that the forecast precision of w-SSVR is better than that of standard SVR. The MAPE of it is only 0.298%, showing the forecast values fit the actual ones very well and this solution is feasible. Last, this thesis tries to optimize the structured variables. The evaluating indicators show that the effect of optimization is good. It not only solves the problems, but also improve the forecast precision further.