The need for disruption in the credit ratings landscape : a model for machine learning computed credit ratings.

Abstract

I present the results from the research on the topics of (1) credit ratings, which are usually provided by credit rating agencies, and (2) Artificial Intelligence and Machine Learning as a form of solving classification tasks, such as credit ratings, without the involvement of human experts. My research problem is stated as follows: to improve the solutions for the credit rating problem introduced by other credit rating agencies, I propose a rating system in the form of an expert system. Then I show that this system is more efficient than traditional rating systems on different hold-out samples of large-scale, multi-period data for public nonfinancial corporate entities worldwide, and with respect to different forecasting horizons. I show that my rating system, which is based on an ensemble machine learning method, specifically Gradient Boosted Decision Trees, when applied to the rating process, outperforms incumbent rating systems on the accuracy-stability scale measured by a compound metric Index of the Quality of Ratings, which I develop and introduce.

In the course of the research in addition to the topic of rating performance evaluation, I have included the comparison of market-implied ratings with fundamental ratings, ratings forecasting and replication, mapping of ratings of different providers to the universal scale, financial effects of qualitative ratings for the investors, the stability of ratings, and the cyclical effects of ratings. The novelty is in the amount of data that I used, including the number and diversification of the rated entities, also in the number of other rating providers involved in performance comparison tests and the number of optional models built and tested. I have shown performance results for different forecasting horizons. The complexity of the proposed model, its iterative revisions throughout the estimation periods, as well as mapping of ratings directly through the default ratios, also mark out my research.

The significance of the research is in showing a more reliable, hi-tech, cost- and timeeffective solution for the problem of credit risk assessment for financial markets participants, who now rely upon the opinion of credit rating agencies. The key output of the research is therefore to re-imagine the credit ratings according to modern advances in finance, datascience, information technology and software. The results of my analysis can be used as a starting point or proxy for choosing the optimal rating agency for investor’s needs, as a stepby- step manual to develop a rating system, as a benchmark for the regulation of rating agencies, or when discussing the quality of ratings in academic and financial papers.