Closing the credit desert: a fairness-constrained geospatial AI framework for predicting and marketing equitable mortgage access in underserved American communities

Geographic credit exclusion – the measurable concentration of mortgage denial in minority-concentrated, low-to-moderate income census tracts represents a structural impediment to household wealth formation in the United States that FICO-based underwriting frameworks were not designed to correct. This study introduces the Geospatial AI Credit Risk and Access (GAICRA) framework, integrating a four-component Geospatial Credit Exclusion Index (GCEI), a fairness-constrained gradient-boosted classifier (FC-CRC), and a Spatial Marketing Allocation Model (SMAM) into a unified, publicly reproducible pipeline for identifying credit deserts and improving mortgage access for thin-file and immigrant borrowers. Using the 2022 Home Mortgage Disclosure Act Loan Application Register for eight states (n = 1,578,187 applications; 37,546 census tracts), supplemented by six additional federal datasets, spatial autocorrelation analysis yields a Global Moran I of 0.436 (p < 0.001), confirming that credit exclusion clusters geographically rather than distributing at random. Local Indicators of Spatial Association identify 4,864 High-High core credit desert tracts. The FC-CRC, trained under a three-way holdout protocol with fairness sample reweighting and separate isotonic calibration, achieves an area under the receiver operating characteristic curve of 0.736 while satisfying the Equal Credit Opportunity Act four-fifths disparate impact standard (ratio = 0.9633). Denied-class recall reaches 43.9% at the F1-optimized threshold of 0.785, a 78% improvement over the unconstrained baseline. A Marketing Priority Score derived from GCEI-Supply and model propensity output provides Community Development Financial Institutions with a quantitative framework for targeting outreach budgets toward the census tracts that combine the highest exclusion severity with the largest pool of approval-eligible-but-currently-denied applicants.