Elements in the Design of an Early Warning System for Sovereign Default

This paper utilizes two different classification techniques to explore issues in the development of an early warning system for sovereign default. Specifically, the paper develops K-means clustering and logit models to illustrate how the optimal choice of parameters, such as assignment rule of fitted observations to binary groups depend on the decision-makers' preferences. It proposes optimization approaches to tailor these parameters to the decision-maker's loss-function and degree of risk-aversion towards unpredicted defaults. The paper also investigates the potential benefits of combining the optimal forecasts from three methods: logit based on objective macroeconomic variables, logit based on judgmental bankers' credit ratings and non-parametric K-means clustering using both objective and judgmental variables. Unlike continuous-variable forecasts, combining forecasts of discrete-variables requires different techniques based on logit regression or voting rules. In this context, the benefit from combination is not as clear-cut, since the expected loss is not directly related to the error variance. We find that the forecast combining approach can also be chosen optimally to account for the decision-makers' loss-function and risk-aversion