A Multimodal Artificial Intelligence Framework for Intelligent Mortgage Risk Assessment Using Underwriting and Loan-Performance Information
Impact in Computics · 07.11.2026 · 2 · 1-23
DOI: 10.65500/computics-2026-009 ↗Kurz gefasst
Fachartikel in der Zeitschrift Impact in Computics, online erschienen am 07.11.2026. Verfasst von Amit Kurchania. Die Arbeit nennt 38 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Mortgage risk models often rely solely on underwriting variables, with unclear targets and evaluation protocols. This study proposes a multimodal artificial intelligence framework using separate encoders for financial and loan-performance features from the same mortgage dataset, adaptive fusion and SHAP explanations. Experiments use Fannie Mae Single-Family Loan Performance Data (2007Q1; 206,914 loans; 4,695 defaults, 2.27%). Default is defined as 90-day delinquency or credit-event exit within 12 months after loan age 12 months, excluding already delinquent loans and non-adverse exits. Evaluation includes stratified 70/15/15 splitting, train-only preprocessing, 10 repeated splits with overlap-corrected inference, loan-disjoint later-age testing and external validation on 377,697 loans from 2018Q1 (0.41% defaults), approximately eleven years later. Compared with Logistic Regression, Random Forest, XGBoost and MLP, the framework achieves ROC-AUC 0.840 ± 0.008 and PR-AUC 0.147 ± 0.011 (random baseline 0.023), with corrected 95% confidence intervals of 0.831–0.850 and 0.133–0.160, respectively. ROC-AUC reaches 0.791 on later-age testing and 0.824 externally, below Random Forest (0.839). Multimodal learning outperforms financial-only (0.797) and loan-performance-only (0.747) models. However, corrected comparisons show adaptive fusion is statistically comparable to concatenation, fixed-weight fusion, Logistic Regression, XGBoost and MLP, outperforming only Random Forest in ROC-AUC on the random split. Class balancing and prevalence shifts require probability recalibration. SHAP highlights FICO score, state, loan purpose and combined LTV. Ethical handling, transparency and human oversight remain essential.
Zitieren
APA
Kurchania, A. (2026). A Multimodal Artificial Intelligence Framework for Intelligent Mortgage Risk Assessment Using Underwriting and Loan-Performance Information. Impact in Computics, 2, 1-23. https://doi.org/10.65500/computics-2026-009
Vancouver
Kurchania A. A Multimodal Artificial Intelligence Framework for Intelligent Mortgage Risk Assessment Using Underwriting and Loan-Performance Information. Impact in Computics. 2026;2:1-23. doi:10.65500/computics-2026-009
BibTeX
@article{Kurchania2026,
author = {Kurchania, Amit},
title = {A Multimodal Artificial Intelligence Framework for Intelligent Mortgage Risk Assessment Using Underwriting and Loan-Performance Information},
journal = {Impact in Computics},
year = {2026},
volume = {2},
pages = {1--23},
doi = {10.65500/computics-2026-009}
}
EUVIMED