Studying the Replicability of Aggregate External Credit Assessments Using Public Information

In this paper, we examine whether an aggregate rating can be accurately predicted with publicly available information about a company's individual characteristics. We propose an algorithm that shows how efficient and replicable an arbitrary aggregate rating is respectively to the widely used credit risk models and to what extent an aggregate rating can be extrapolated to the non-rated companies as a valid indicator of their credit risk. Using this algorithm, we empirically study the aggregate ratings constructed as a consensus of ratings assigned by seven credit rating agencies for Russian banks on a national scale and compare it with several alternatives and proxies based on the publicly available characteristics of those banks. We measure how well the aggregate (consensus) rating and the proxies are agreed in terms of ordering banks by their credit quality and predicting defaults over a one-year horizon. We show that the aggregate (consensus) rating is comparable to a standard logit default model in terms of discriminatory power, but for ordering, the former is in low agreement with the latter. We also found that using models for predicting initial credit ratings allows the building of a proxy that is in high agreement with the original aggregate rating, but the original aggregate rating outperforms the proxy in terms of discriminatory power. It was also found that greater agreement between the original aggregated rating and the proxy can be achieved on a subsample of investment grade ratings