Prethodno priopćenje

https://doi.org/doi.org/10.32910/ep.76.1.2

A Bayesian approach to logistic regression for bankruptcy prediction of trading companies under uncertainty conditions

Josip Arnerić
Matteo Moćan

Sažetak

This article provides solutions to common methodological issues in studies dealing with bankruptcy prediction of trading companies, such as the selection of a homogeneous random sample, the choice of financial and non-financial indicators relevant for bankruptcy prediction, determination of an optimal probability threshold for classification, the specification of prior distributions for Bayesian logistic regression applicability, and thus it adds value to the existing literature. Considering the necessity of bankruptcy prediction as the most effective way of its prevention, particularly during crisis periods, requirements for improving forecasting models are imposed. In this context, it was examined not only which indicators are significant for bankruptcy prediction in the wholesale and retail trade sector but also whether the Bayesian approach contributes to the predictive ability of logistic regression during uncertain periods, such as the COVID pandemic when the same method was employed. The adequacy of the Cauchy prior for the constant term is proved, while for the remaining coefficients the Student’s t-prior with a standard deviation of 2.5 on the logarithmic odds ratio scale is more applicable. The short-term assets turnover is identified as the indicator that best discriminates bankrupt from non-bankrupt companies, although it is concurrently the most uncertain. Nevertheless, the EBIT margin exhibits the lowest discriminatory power, but compared to the other variables is the most certain, while indebtedness and current liquidity are moderately uncertain variables. The predictive accuracy of Bayesian logistic regression of 90.22% in 2021 is documented by cross-validation, whereby 99.61% of bankrupt trading companies are correctly classified, with a lower cut-off probability compared to standard logistic regression. Adequately specified weakly informative prior’s, which are crucial for posterior analysis of the parameters, are the main contribution of the research, and consequently the determination of uncertainty level for each bankruptcy indicator individually.

Ključne riječi

Bayesian logistic regression; bankruptcy prediction; financial and non-financial indicators; wholesale and retail trade, prior distributions of parameters

Hrčak ID:

329121

URI

https://hrcak.srce.hr/329121

Datum izdavanja:

14.3.2025.

Podaci na drugim jezicima: hrvatski

Posjeta: 0 *