Penalties and Premiums in Sovereign Credit Ratings

Abstract

Credit rating agencies collapse high-dimensional borrower characteristics into summaries of creditworthiness, facilitating capital flows. But biases embedded in rating algorithms may lead to misallocation. We test for bias in sovereign credit ratings across a wide array of borrower-country characteristics, training machine learning models to estimate ratings as a function of countries’ observable economic, political, and borrower history fundamentals. Even after accounting for these fundamentals, ratings agencies tend to favor the “clubs” of the Western world, namely the members of the G7, EU, and OECD, while penalizing emerging Latin American and Asian nations. Using data on sovereign bond issues, we find that these penalties and premiums increase coupon spreads between the G7 (the most overrated) and Southeast Asia (the most underrated) by 62.7 basis points. We show that it is possible to earn risk-free excess returns by using our algorithm to construct an unbiased portfolio of investment-grade bonds, suggesting persistent mispricing.

Keywords: sovereign debt, credit ratings, algorithmic bias, international finance, machine learning