OptiGeoRisk: A Spatially Localized and Optimized Geo-XGBoost Framework for Credit Default Risk Classification

Credit default risk prediction is important for financial institutions in geographically diverse, collateral-based lending, where early identification of risky borrowers supports preventive monitoring. Most credit scoring models treat borrower records as independent observations and rarely capture local spatial heterogeneity. This study proposes OptiGeoRisk, a spatially localized Geo-XGBoost framework for imbalanced credit default classification, using 76,742 loan transactions from West Java, Indonesia. OptiGeoRisk treats geographic coordinates as a local learning mechanism, combining haversine-based nearest-neighbor selection, adaptive neighborhood size and kernel bandwidth, and global-local probability fusion, and is evaluated against XGBoost Non-Spatial and XGBoost Spatial Feature baselines using PR-AUC, categorical and Top-K accuracy, default recall, precision, and false-positive volume. XGBoost Spatial Feature obtains the best PR-AUC, while OptiGeoRisk achieves the highest default recall, trading global ranking quality for localized early-warning sensitivity. Bootstrap resampling shows this recall advantage is statistically robust, while OptiGeoRisk’s PR-AUC is statistically indistinguishable from XGBoost Non-Spatial and significantly lower than XGBoost Spatial Feature; refitting under 10 random seeds confirms this pattern is stable, and a cross-region check in DKI Jakarta shows the PR-AUC gain over a no-skill baseline holds outside West Java, though the recall advantage does not. A leakage-robust check across borrower-grouped, spatially blocked, and temporal splits shows the recall advantage holds under random and borrower-grouped partitioning but weakens when whole areas or future periods are held out, a limitation traced mainly to adaptive neighborhood selection and imbalance handling rather than the spatial kernel. These findings position OptiGeoRisk as a localized early-warning layer complementing a well-tuned global model.

@article{Prastian2026,
  title     = {OptiGeoRisk: A Spatially Localized and Optimized Geo-XGBoost Framework for Credit Default Risk Classification},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {9},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Claher Prastian and Alexander Agung Santoso Gunawan},
  doi       = {10.14569/IJACSA.2026.0170907},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170907}
}