Predicting LDC debt rescheduling: performance evaluation of OLS, logit, and neural network models

Abstract

Empirical studies in the area of sovereign debt have used statistical models singularly to predict the probability of debt rescheduling. Unfortunately, researchers have made few efforts to test the reliability of these model predictions or to identify a superior prediction model among competing models. This paper tested neural network, OLS, and logit models' predictive abilities regarding debt rescheduling of less developed countries (LDC). All models predicted well out-of-sample. The results demonstrated a consistent performance of all models, indicating that researchers and practitioners can rely on neural networks or on the traditional statistical models to give useful predictions.

Understanding and predicting sovereign debt rescheduling: a comparison of the areas under receiver operating characteristic curves

“…This finding is in line with what was found in Barney and Alse (2001), who rely on comparing mean squared errors across models. However, the MSE criterion sometimes could be misleading as to what model to use.…”

“…For someone that had interpretation needs, the tree-based models and the MARS additive model would be preferred to the neural network. Barney and Alse (2001) find that their models are equally reliable. According to Table XV, only seven out of 21 comparisons between models indicate something other than equal reliability of models.…”

“…In a recent paper, Barney and Alse (2001) address the issues of reliability and comparability among competing models only in regard to their debt-rescheduling predictive power. They provide a test 1 The principal problem that the researcher faces is the specification of the relationship between the response variable, Y, and the explanatory factors x = {x 1 , .…”

“…In a recent paper, Barney and Alse (2001) address the issues of reliability and comparability among competing models only in regard to their debt-rescheduling predictive power. They provide a test (2) Equation (2) does not generate a unique solution (Friedman, 1994).…”

“…The issue of interpretability of models for debt rescheduling is not addressed by Barney and Alse (2001). In this paper, the interpretability is addressed in the four models used here.…”