2013 Volume 23 Issue 1 Pages 70-91
A large-scale experiment using a support vector regression (SVR) to analyze the relationship between sovereign bond ratings of 158 countries and 44 kinds of economic and political indices was carried out to construct a model for reproducing ratings of various countries just from publicly open numerical information. Effective indices were selected by using the sensitivity analysis method, and the SVR model was optimized. The ratings of various countries could satisfactorily be reproduced from published indices with the root-mean-square error of 0.070 and the coefficient of determination of 0.925. The contribution of political indices is more significant than economical ones, suggesting that rating companies consider political factors important in rating sovereign bonds. Foreign rating companies severely judge the Japan's sovereign bonds by critically estimating the accumulated debt while domestic rating companies judge high compared with index data. The relationship between ratings and indices is highly nonlinear, therefore the combination of SVR and sensitivity analysis is indispensable for constructing a model to reproduce sovereign bond ratings.