Validity of traditional bankruptcy prediction models during crisis periods

Details

Supervisors

Thewissen, James

Faculty

Louvain School of Management

Degree label

Master [120] en sciences de gestion, à finalité spécialisée

Abstract

enThis thesis evaluates the out-of-sample predictive validity of three traditional accounting-based bankruptcy prediction models Altman's Z-score (1968), Altman's Z''-score (1983), and Ohlson's O-score (1980), on a sample of 104,367 U.S. public firm-year observations drawn from Compustat over 1998-2022, including 377 bankrupt firms across three macroeconomic crisis regimes and ten Fama-French industry groups, to determine their robustness over time. All three models are applied in their original form and re-estimated on modern training data, with classification performance evaluated on a common holdout sample of 148 bankrupt and 2,960 non-bankrupt firm-years. The results indicate that all three models report substantial performance degradation when applying their original published coefficients, driven by structural changes in firm financial profiles since their original estimation periods. Re-estimation recovers meaningful predictive validity for Ohlson’s O-score and Altman Z’’-score model specifications, achieving combined errors of 30.4% at C* = 0.051 and 35.2% at C* = 0.107, respectively, but fails to do so for the original Altman Z-score. Coefficient analysis reveals a structural shift in the determinants of bankruptcy with leverage losing its predictive weight while the importance of sustained operating losses increases nearly sixfold. Counterintuitively, model performance does not deteriorate considerably during the Global Financial Crisis or Dot-com bust but varies significantly across Fama-French industry groups with the strongest results observed for Business Equipment at a combined error rate of 25.8%. The findings indicate that the Ohlson variable set is the most temporally robust of the three when re-estimating, that cutoff recalibration offers the most accessible performance gain for practitioners, and that industry heterogeneity underscores the need for sector-specific and dynamically updated prediction frameworks.