Same Split, Same Champion, Uneven Errors: A Pre-registered R Rebuild of a Credit-Risk Champion That Matched on Discrimination, a GAM Refuted by 0.0009 AUC, and an Equal-Opportunity Gap of 39.16 Points on Education

A published model that cannot be rebuilt in a second language, on the identical data split, has not been fully reported: some part of it lives in the original code rather than in the paper. This study is that rebuild. The Python credit-risk champion of the author's credit-risk-lab study (HistGradientBoosting, test AUC 0.7810 on the UCI Taiwan credit card portfolio, 30,000 accounts) was rebuilt natively in R (tidymodels: glmnet ridge logistic regression, gradient boosting through the xgboost engine, and an mgcv GAM arm), under hypotheses, a threshold, and verdict rules frozen in a pre-registration committed before any model was fit. The locked Python split (stratified 70/30, seed 42) was reimplemented in pure R and proven identical to the Python split by hash, not by summary statistics: the SHA-256 of the sorted test positions matches the Python reference exactly (test n 9,000, event rate 0.22122). H1 was CONFIRMED: the R gradient boosting arm scored AUC 0.78237, a difference of 0.00137 against the champion, inside the frozen 0.01 band, with KS 0.42429 against 0.4196 and Brier 0.13502 against 0.1355. H2 was REFUTED by 0.00090 AUC: the GAM, the fully interpretable arm (deviance explained 16.7 percent, full smooth-term table published), scored AUC 0.76147, giving back 0.02090 against a frozen allowance of 0.02. H3 was CONFIRMED: at the frozen champion threshold of 0.2293, the rebuilt model's equal-opportunity gap on education is 39.16 percentage points (the small "others" group, n 120 with 8 defaults, at TPR 0.2500 against the university group's 0.6416), the age gap is 5.78 points on large groups, and the sex gap is 1.21 points; predictive-parity gaps are 19.46, 5.33, and 0.21 points. The scope is one public 2005 Taiwan portfolio and one locked split, and the claims are method claims: the champion's discrimination survives a change of language and toolkit, the interpretability price of the GAM is real and slightly larger than the tolerance set in advance, and the rebuilt model's errors fall unevenly on protected groups at the operating point it inherited.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-18
DOI
https://doi.org/10.5281/zenodo.24328872
Primary Topic
Financial Distress and Bankruptcy Prediction
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Same Split, Same Champion, Uneven Errors: A Pre-registered R Rebuild of a Credit-Risk Champion That Matched on Discrimination, a GAM Refuted by 0.0009 AUC, and an Equal-Opportunity Gap of 39.16 Points on Education

Jesun Ahmad Ushno
1 links
Zenodo (CERN European Organization for Nuclear Research)
Financial Distress and Bankruptcy Prediction
preprint

Same Split, Same Champion, Uneven Errors: A Pre-registered R Rebuild of a Credit-Risk Champion That Matched on Discrimination, a GAM Refuted by 0.0009 AUC, and an Equal-Opportunity Gap of 39.16 Points on Education

Jesun Ahmad Ushno
preprint en

Abstract

A published model that cannot be rebuilt in a second language, on the identical data split, has not been fully reported: some part of it lives in the original code rather than in the paper. This study is that rebuild. The Python credit-risk champion of the author's credit-risk-lab study (HistGradientBoosting, test AUC 0.7810 on the UCI Taiwan credit card portfolio, 30,000 accounts) was rebuilt natively in R (tidymodels: glmnet ridge logistic regression, gradient boosting through the xgboost engine, and an mgcv GAM arm), under hypotheses, a threshold, and verdict rules frozen in a pre-registration committed before any model was fit. The locked Python split (stratified 70/30, seed 42) was reimplemented in pure R and proven identical to the Python split by hash, not by summary statistics: the SHA-256 of the sorted test positions matches the Python reference exactly (test n 9,000, event rate 0.22122). H1 was CONFIRMED: the R gradient boosting arm scored AUC 0.78237, a difference of 0.00137 against the champion, inside the frozen 0.01 band, with KS 0.42429 against 0.4196 and Brier 0.13502 against 0.1355. H2 was REFUTED by 0.00090 AUC: the GAM, the fully interpretable arm (deviance explained 16.7 percent, full smooth-term table published), scored AUC 0.76147, giving back 0.02090 against a frozen allowance of 0.02. H3 was CONFIRMED: at the frozen champion threshold of 0.2293, the rebuilt model's equal-opportunity gap on education is 39.16 percentage points (the small "others" group, n 120 with 8 defaults, at TPR 0.2500 against the university group's 0.6416), the age gap is 5.78 points on large groups, and the sex gap is 1.21 points; predictive-parity gaps are 19.46, 5.33, and 0.21 points. The scope is one public 2005 Taiwan portfolio and one locked split, and the claims are method claims: the champion's discrimination survives a change of language and toolkit, the interpretability price of the GAM is real and slightly larger than the tolerance set in advance, and the rebuilt model's errors fall unevenly on protected groups at the operating point it inherited.

Zenodo (CERN European Organization for Nuclear Research)
Reduced inequalities
Financial Distress and Bankruptcy Prediction
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon