In this dissertation a new method for the prediction of the Brazilian Term Structure of Interest Rates - Brazilian ETTJ - known as Support Vector Regression is investigated. This is compared with the traditional methods used in this set up, such as VAR models (Vector Autoregressive) and ECM (Error Correction Models). Besides the interest rates, some macroeconomic variables are also used, as it was suggested in a work from Evans and Marshall(1998) and verified for brazilian economy in a work from Fukuda, Vereda and Lopes (2006), the inclusion of macroeconomic variables can improve the prediction of the interest rates in long term forecasts. The experiment show some improvements in using SVR in the long term in relation to the traditional methods mentioned, acting like a realy good predictor of the direction of the interest rates along the short and long term forecasts. To make these assertions, we make use of some tests like the root mean squared error, mean absolute error, directional symmetry and weighted directional symmetry, Correct Up trend and Corret Down trend besides Theil U test, which uses the root mean squared error to verify if there is some significant improvement between two models. As there is not a structured way to choose the free parameters of SVR, a function in the R software was used in order to make a grid search over a supplied parameter ranges. The analysis of the results demonstrate that SVR is a promising technique to prediction of interest rates, suggestions are also made in order to get better the choices of the free SVR parameters once they are powerful means of regularization and adaptation to the noise in the data.