Validating Mortgage Default Models Through a Regime Change: Evidence from 10.8 Million Freddie Mac Loans, 2020 to 2026

Credit default models are usually validated on data from one stable period but used through regime changes. This paper validates default models on 10.8 million Freddie Mac mortgages with first payments from 2020 to mid-2024 and performance through June 2026, a period that spans COVID-19 forbearance, a refinancing boom and a rise in note rates from 2.8% to 7.3%. Default is 90 or more days of delinquency or a credit event within 24 months, excluding delinquency that began under forbearance. All data-dependent steps were fitted inside training folds, and models built on 2020 to 2022 loans were tested on a held-out sample and on all 1.23 million loans from 2023 and 2024. An ensemble of logistic regression and gradient boosting ranked risk best, with a ROC AUC of 0.807 in time (95% BCa interval 0.799 to 0.816) and 0.815 out of time (0.811 to 0.817), a small but significant gain over logistic regression. Repeated stratified cross-validation gave the most stable estimates, LOOCV was twice as biased and 400 times slower, and none of eight resampling methods improved ranking or cost. Predicted risk tracked a 2.6-fold rise in default rates because the riskier borrower mix was visible at origination, yet a fixed review cutoff flagged 24% of later loans instead of 5%, and about four in ten flagged loans were unstable across bootstrap refits. Early prepayment explained much of the calibration error by vintage, and only 5% of defaults ended in a credit event, so realized losses did not justify declining applications. The findings support capacity based review thresholds, population stability monitoring and competing risks models of default and prepayment. Declaration of generative AI use. During the preparation of this work the author used Claude (Anthropic) to write and run the analysis code, to produce the tables and figures, and to draft the text of the manuscript, under the author's direction. The author specified the scope of the study, obtained the data, and reviewed, edited and verified the analyses, interpretations and text, and takes full responsibility for the content of this paper. Data. Freddie Mac Single-Family Loan-Level Dataset, Release 48, used under Freddie Mac's terms for academic research. Freddie Mac provides the data as is and is not responsible for the analysis or conclusions of this paper. JEL classification: C52, C53, C55, G21.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-18
DOI
https://doi.org/10.5281/zenodo.24358933
Primary Topic
Credit Risk and Financial Regulations
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Validating Mortgage Default Models Through a Regime Change: Evidence from 10.8 Million Freddie Mac Loans, 2020 to 2026

Daniel (Zhenxiong) Yang
Zenodo (CERN European Organization for Nuclear Research)
Credit Risk and Financial Regulations
preprint

Validating Mortgage Default Models Through a Regime Change: Evidence from 10.8 Million Freddie Mac Loans, 2020 to 2026

Daniel (Zhenxiong) Yang
preprint en

Abstract

Credit default models are usually validated on data from one stable period but used through regime changes. This paper validates default models on 10.8 million Freddie Mac mortgages with first payments from 2020 to mid-2024 and performance through June 2026, a period that spans COVID-19 forbearance, a refinancing boom and a rise in note rates from 2.8% to 7.3%. Default is 90 or more days of delinquency or a credit event within 24 months, excluding delinquency that began under forbearance. All data-dependent steps were fitted inside training folds, and models built on 2020 to 2022 loans were tested on a held-out sample and on all 1.23 million loans from 2023 and 2024. An ensemble of logistic regression and gradient boosting ranked risk best, with a ROC AUC of 0.807 in time (95% BCa interval 0.799 to 0.816) and 0.815 out of time (0.811 to 0.817), a small but significant gain over logistic regression. Repeated stratified cross-validation gave the most stable estimates, LOOCV was twice as biased and 400 times slower, and none of eight resampling methods improved ranking or cost. Predicted risk tracked a 2.6-fold rise in default rates because the riskier borrower mix was visible at origination, yet a fixed review cutoff flagged 24% of later loans instead of 5%, and about four in ten flagged loans were unstable across bootstrap refits. Early prepayment explained much of the calibration error by vintage, and only 5% of defaults ended in a credit event, so realized losses did not justify declining applications. The findings support capacity based review thresholds, population stability monitoring and competing risks models of default and prepayment. Declaration of generative AI use. During the preparation of this work the author used Claude (Anthropic) to write and run the analysis code, to produce the tables and figures, and to draft the text of the manuscript, under the author's direction. The author specified the scope of the study, obtained the data, and reviewed, edited and verified the analyses, interpretations and text, and takes full responsibility for the content of this paper. Data. Freddie Mac Single-Family Loan-Level Dataset, Release 48, used under Freddie Mac's terms for academic research. Freddie Mac provides the data as is and is not responsible for the analysis or conclusions of this paper. JEL classification: C52, C53, C55, G21.

Zenodo (CERN European Organization for Nuclear Research)
University of the Cumberlands (US)
Credit Risk and Financial Regulations
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