Size Matters: A Lag‐Aware Evaluation of Bankruptcy Prediction Models in Ecuador (2021–2024)

ABSTRACT This study evaluates five accounting‐based bankruptcy prediction models—Altman Z′′ (1998), Pascale (1988), Springate (1978), Zmijewski (1984) and CA‐Score (Legault, 1987)—in Ecuador's corporate sector during the postpandemic period 2021–2024, using an Ecuadorian administrative firm‐year panel ( n = 324,252; 110,404 firms). The design departs from prior validation work in the following five ways: continuous winsorised scores replace dichotomised flags, with thresholds re‐estimated via Youden's J per model and size stratum; performance is reported through prevalence‐adjusted metrics (Matthews correlation, F1, Bayesian PPV and decision‐curve net benefit) better suited to the sample's 1:42 class imbalance; results are tested across three definitions of legal failure; a lag‐aware robustness check accounts for the 2‐ to‐3‐year administrative dissolution lag under Article 20 of the Companies Law; and an XGBoost benchmark (70/30 stratified split) bounds the out‐of‐sample gap. Our results suggest that model validity is size‐conditional: Altman rises monotonically (AUC 0.553 → 0.657; z = +3.84), Springate declines (0.623 → 0.489; z = −5.14) and Pascale falls below random among large firms (0.464). Continuous‐score evaluation reorders the field: Springate (0.624) > Pascale (0.594) > CA‐Score (0.588) > Altman (0.562) > Zmijewski (0.553). Even at recalibrated thresholds, Bayesian PPV stays below 5.5%, reframing these flags as portfolio‐level tools rather than firm‐level verdicts. The Institutional Friction Index reveals a 7 × −73 × wedge between predicted distress and legal liquidation, decomposed into real‐recovery (~10%), administrative‐lag (~50%) and panel‐composition (~40%) components. XGBoost reaches out‐of‐sample AUC 0.796 (vs. 0.694 for the five scores), showing traditional models leave roughly one‐third of attainable performance untapped. The findings support a segmented, decision‐calibrated, lag‐aware early‐warning regime over single‐threshold gatekeeping.