Bankruptcy prediction with fractional polynomial transformation of financial ratios

We show that simple nonlinear transformations of financial ratios, within a multivariate fractional polynomial approach, yield substantial improvements in bankruptcy prediction. The approach selects optimal power functions balancing parsimony and complexity. Focusing on a dataset comprising of non-financial firms, we develop a parsimonious nonlinear logit model with minimal parameter specification and clear interpretability, outperforming linear logit models. The model improves the in-sample fit, while out-of-sample it significantly reduces costly misclassification errors and improves discriminatory power. Similar insights are obtained when applying fractional polynomials on a secondary dataset consisting of banking firms. Interestingly, the fractional polynomial model compares favourably with other nonlinear models. By simulating a competitive loan market, we demonstrate that the bank using the fractional polynomial model builds a higher-quality loan portfolio, resulting in superior risk-adjusted profitability compared to banks employing alternative models.