On Unbalanced Sampling in Bankruptcy Prediction

Abstract

The paper discusses methodological topics of bankruptcy prediction modelling-unbalanced sampling, sample bias, and unbiased predictions of bankruptcy. Bankruptcy models are typically estimated with the use of non-random samples, which creates sample choice biases. We consider two types of unbalanced samples: (a) when bankrupt and non-bankrupt companies enter the sample in unequal numbers; and (b) when sample composition allows for different ratios of bankrupt and non-bankrupt companies than those in the population. An imbalance of type (b), being more general, is examined in several sections of the paper. We offer an extended view of the relationship between the biased and unbiased estimated probabilities of bankruptcy-probability of default (PD). A common error in applications is neglecting the possibility of calibrating the PD obtained from a bankruptcy model to the unbiased PD that is population adjusted. We show that Skogsviks' formula of 2013 coincides with prior correction known for the logit model. This, together with solutions for other binomial models, serves as practical advice for obtaining the calibration of unbiased PDs from popular bankruptcy models. In the final section, we explore sample bias effects on classification. Zhou, Ligang. 2013. Performance of corporate bankruptcy prediction models on imbalanced data set: The effect of sampling methods. Knowledge-Based Systems 41: 16-25.[CrossRef] Zmijewski, Mark E. 1984. Methodological issues related to the estimation of financial distress prediction models.

Women on Boards and Firm Performance: A Microeconometric Search for a Connection

“…When samples are unbalanced, Cramer (1999) advocates the use of a cut-off point α equal to the proportion of ones in the sample. In effect, the success rates for y i = 1 and y i = 0 are better spread than for the typical cut-off point of 0.5 (Gruszczyński 2019).…”

Machine Learning for Zombie Hunting. Firms' Failures and Financial Constraints.

“…To clarify the nature of the problem, if we applied a methodology that predicted all the firms to fail in our data, we would still get an accuracy measure for correctly classified observations over total observations of 94.3%. To deal with similar unbalanced outcomes, one could either under-sample the label that accounts for the largest majority of observations, or oversample observations for the other label, as suggested in other applications (Gruszczyński, 2019;Zhou, 2013). In our case, we argue that any re-sampling method is sub-optimal because we know from Section 2, Table 2, and Appendix Table A2 that patterns of missing values are positively correlated with the outcome.…”

Urban–Rural Gradients Predict Educational Gaps: Evidence from a Machine Learning Approach Involving Academic Performance and Impervious Surfaces in Ecuador

“…This exchange of student data among neighboring cases blurred the results, especially in areas where large distances were required (see Figure 3). Nevertheless, this operation allowed us to avoid unbalanced sampling [77] and to prepare data for feature selection and model training. Our results indicated that the selected features of the models using Boruta resulted in RF predictions with overall scores of around 0.83-0.84% and Kappa values of 0.65-0.67%.…”