Creative Commons

This dissertation was written as part of the MSc in Banking and Finance at the International Hellenic University. The main purpose of this study is to present an empirical model designed to forecast banks’ credit ratings using information from their financial statements. For this reason, we have used ratings provided by Fitch in 2012. The sample consists of 92 US banks with their financial statements from 2008 to 2011. Using the same financial data and for the same time frame as (Go gas, Papadimitriou and Agrapetidou 2014) in their study, we used machine learning (ML) models in order to examine whether they are more efficient than classical econometric techniques that were used on their research . For this purpose using the same data as (Gogas, Papadimitriou and Agrapetidou 2014) and following their classification we trained the data both with linear and non - linear support vector machines in order to examine whether the prediction accuracy is higher with support vector machines rather than with ordered probit model . According to the simulation results ,in the optimal group of regressors, both linear and non - linear (RBF kernel) Support Vector Machines , can predict more accurately credit ratings with a 84 . 06 percent accuracy for the linear SVM which is slightly higher than the 83 . 7 0 percent accuracy achieved by the ordered probit model of (Gogas, Papadimitriou and Agrapetidou 2014) . On the other hand, non - linear SVMs can predict much more accurately than ordered probit models with 99 . 64 percent prediction accuracy. In other words, using either Linear Support Vector Machine or non - linear RBF kernel we can predict credit ratings more accurately than Ordered Probit model.