IN SEARCH OF A PARSIMONIOUS BANKRUPTCY MODEL FOR PRIVATE FIRMS AND THE COST TO LENDERS

Sammanfattning: This paper comprehensively reviews in excess of 200 financial and non-financial covariates in search of a parsimonious bankruptcy prediction model for the private market. Financial and real-estate companies aside, the entire population of Swedish independent, limited liability companies are examined between 1998-2017 corresponding to 245,844 unique companies and 55,411 corporate default events. Predictors are atheoretically selected for one- to five-year forecasts and fitted as discrete-time duration-dependent hazard models with and without a frailty term. Each of these is cross-validated and tested against five competing covariate specifications: Altman's (1968) Z-score, Zmijewski (1984), Shumway (2001), Altman and Sabato (2007) and Dakovic et al. (2010). For the vast majority of prediction horizons, estimation windows and hold-out tests, we significantly outperform these contesting models along dimensions of model power, calibration and early-retrieval performance as measured by principally ROC curves, PR curves and their integrals. From our covariate vectors, we find evidence of intra-industry contagion at work for longer look-ahead predictions, suggest employees may be an underused denomination of company size and discover untaxed reserves as a highly promising indicator of bankruptcy in the context of Swedish Private Company Law. Finally, we gauge the practical merits of our empirically derived covariate specifications by backtesting all six one-year predictive models on simulated (equally-weighted) and actual private loans (contained in the data) during 2014-2016. The lender's perspective is adopted using the Basel III F-IRB approach. In both settings, the bank that assumes our bankruptcy hazard model makes considerably higher returns on risk-weighted assets than any of its competition.