Using Meta Learning Methods to Forecast Sub-Sovereign Credit Ratings

Abstract

This paper mainly analyses the forecasting of sub-sovereign credit ratings using machine learning methods in the non-US, Europe and other regional and sub-sovereign ratings. Specific focus is based on developing an accurate forecasting model based on machine learning. The forecasting accuracy was examined on two forecasting horizons, one and two years ahead. The study was designed to determine the cost sensitivity of various machine learning methods and to develop an accurate decision-support system that minimizes the cost of credit rating classification for sub-sovereign entities across countries and world regions. Each side of the economic, financial and debt and budget, revenues and expenditures were considered to provide sufficient inputs for the machine learning models. The analyses is to consider the ordinal character of the rating classes, classification cost (cost-sensitivity) which is used as objective function, in assessing credit ratings and evaluating of bonds i.e. regional credit rating modeling. This paper has been able to demonstrate that machine learning models based on current available financial and economic data present accurate classifications of credit ratings. Also the sub-sovereign credit rating forecast signified that the Random Forest and SMO algorithm performed significantly better than the statistical methods. Some practical implications were also provided.