In this study, we examine the predictive performance of a wide class of binary classifiers using a large sample of international credit ratings changes from the period 1983–2013. Using a number of financial, market, corporate governance, macro-economic and other indicators as explanatory variables, we compare classifiers ranging from conventional techniques (such as logit/probit and LDA) to fully nonlinear classifiers, including neural networks, support vector machines and more recent statistical learning techniques such as generalised boosting, AdaBoost and random forests. We find that the newer classifiers significantly outperform all other classifiers on both the cross sectional and longitudinal test samples; and prove remarkably robust to different data structures and assumptions. Simple linear classifiers such as logit/probit and LDA are found nonetheless to predict quite accurately on the test samples, in some cases performing comparably well to more flexible model structures. We conclude that simpler classifiers can be viable alternatives to more sophisticated approaches, particularly if interpretability is an important objective of the modelling exercise. We also suggest effective ways to enhance the predictive performance of many of the binary classifiers examined in this study.

Over the past three decades the literature on financial distress prediction has largely been confined to simple multiple discriminant analysis, binary logistic or probit analysis, or rudimentary multinomial logit models (MNL). There has been a conspicuous absence of modeling innovation in this literature as well as a failure to keep abreast of important methodological developments emerging in other fields of the social sciences. In particular, there has been no recognition of major advances in discrete choice modeling over the last 15 years, which has increasingly relaxed behaviorally questionable assumptions associated with the independently and identically distributed errors (IID) condition and allowed for observed and unobserved heterogeneity. In contrast to standard logit, the mixed logit model fulfils this purpose and provides a superior framework for explanation and prediction. We explain the theoretical and econometric underpinnings of mixed logit and demonstrate its empirical usefulness in the context of a specific but topical area of accounting research: financial distress prediction. Comparisons of model-fits and out-of-sample forecasts indicate that mixed logit outperforms standard logit by significant margins. While mixed logit has valuable applications in financial distress research, its potential usefulness in other areas of accounting research should not be overlooked.