A predictive model of sovereign investment grade using machine learning and natural language processing

Abstract

Rating agencies like Moody’s, Standard and Poor’s and Fitch rate sovereign assets based on mathematical analysis of economic, social and political factors and expert judgment. According to the rating, sovereign can be classified as having investment grade or speculative status. Having an investor grade is important as it reduces the cost of financing and expands the pool of potential investors in an economy. In this paper we want to predict whether a sovereign has investment grade status using macroeconomic variables and text analysis variables obtained from the reports issued by Fitch between 2000 and 2018 using natural language processing techniques. We use logistic regression and a series of alternative machine learning algorithms as k-nearest neighbors, support vector machine, classification and decision trees and random forest. According to our results report’s sentiments, captured by the uncertainty index is statistically significant to predict investment grade status. When comparing the different algorithms random forest has the best predictive performance out of sample when the independent variables are referred to the same year and random forest and k-nearest neighbors have the best predictive performance when the independent variables are referred to one year before in terms of f1-score and recall.