Outcome leakage in firm-failure prediction: evidence from a Romanian registry extract

Firm-failure models are routinely estimated on commercial registry extracts in which accounting histories and the firm’s legal status are both read at extraction. We show that this common design inflates measured discrimination, and we quantify by how much. Using a licensed Wolters Kluwer extract of 29,183 Romanian construction firms, we fix a landmark at the end of 2020 and predict a first entry in the national insolvency bulletin during 2021–2025 among the 14,393 firms carrying no such entry at the landmark. Holding the sample, the events and the estimator fixed, and varying only whether accounting years after the landmark may enter the predictors, ROC-AUC moves from 0.728 to 0.864 and PR-AUC from 0.165 to 0.345; the AUC difference is +0.136 [+0.117, +0.156]. The inflation is concentrated in final-year and trend features, which carry almost no signal at the landmark and much more once the window runs past the exit. The clean model is stable across four landmarks (0.728–0.738). Undated legal records leak through the same channel: a litigation snapshot taken at extraction appears to add +0.104 AUC, which is an upper bound rather than a predictive gain.