Determinants of corporate default: a BMA approach

Abstract

In this article, we aim to identify the main determinants of corporate default by considering Bayesian Model Averaging (BMA) techniques. Our empirical findings suggest that the most robust determinants of firm default are firm-specific variables such as the ratio of working capital to total assets and the SD of the firm stock return. In contrast, aggregate variables do not seem to play a relevant role once firm-specific characteristics (observable and unobservable) and model uncertainty are taken into consideration.

Keywords:

• Bayesian model averaging
• corporate default
• panel data
• credit risk

JEL Classification:

• C11
• G33
• C23

Acknowledgements

The authors thank Cristian Bartolucci, Max Bruche and Joan Llull for their helpful comments and suggestions. Research funding from the Spanish Ministry of Science and Innovation, Consolider Grant CSD2006-00016 is gratefully acknowledged.

Notes

¹ Furthermore, besides appropriate confidence bands (BMA's SE analogues incorporate both the estimated variances in individual models as well as the variance in estimates of the coefficients across different models), BMA also provides a measure of robustness of each variable in predicting firm default, the Posterior Inclusion Probability (PIP) (see Koop (2003) for more details on BMA). Simonian (2011) discusses an alternative Bayesian approach to credit default modelling.

² In the absence of a positive theory of default, we select the logistic function given its popularity in the credit risk literature and its convenience when including firm-specific effects in the model.

³ Bonfim (2009) considers Gaussian firm-specific effects in her empirical model under the assumption that they are uncorrelated with the other right-hand-side variables (i.e. random-effects specification).

⁴ Results based on the normal distribution (i.e. probit specification) are available upon request and very similar to those based on the logit specification. For the case with firm-specific effects, a probit fixed-effects specification is not available given the lack of a sufficient statistic for the firm effects in the probit likelihood function.

⁵ See Koop (2003) for more details on the BMA methodology and Moral-Benito (2012) for its generalization to panel data settings.

⁶ Results based on the probit specification for the model-specific step are in line with those presented in Table 1 and are available upon request.

⁷ Li and Zhao (2006) document variation in default rates across industries.

⁸ As already pointed out in Altman (1968), this ratio implicitly considers the age of a firm since it might take sometime to build up the cumulative profits.

⁹ Estimates based on single model specifications provide evidence in favour of this hypothesis. These results are available upon request and also in the working paper version of this article.