Forecasting Banking Desert Risk Using Branch Level and Merger Signals

In the United States, banking desert status is classified annually from that same year’s branch data, producing a contemporaneous snapshot of current access rather than a forecast of which communities are likely to change status next. This study tests whether a census-tract-level status can be predicted one year ahead, using FDIC Summary of Deposits branch-level deposit records, Federal Reserve Communities desert-status classifications, ACS broadband access estimates, and a branch-level indicator of merger, evaluated through a temporally ordered backtest across five prediction years (2021-2025; 83,848 tracts; 503,088 tract-years). Three models are compared on a three-category outcome (2+Branches, Potential Desert, Banking Desert): naïve persistence, logistic regression, and XGBoost. On the full panel, naïve persistence achieves the highest mean macro-F1 (0.979); among the learned models, XGBoost outperforms logistic regression on every fold (mean macro-F1 0.741 vs 0.558). Restricted to the 1,471 tract-years where status actually changed, this ranking reverses in all ten fold-by-feature-set combinations tested: logistic regression outperforms XGBoost (mean macro-F1 0.306 vs 0.156; mean macro AUC-PR 0.339 vs 0.254), a reversal statistically significant by McNemar’s test (chi-squared = 83.37, p=6.80e-20). A cost-weighted analysis of the transition subset identifies a breakeven point (missed-desert-to-false-alarm cost ratio of 0.71) above which logistic regression is the more defensible operational choice. Removing broadband-related features produces no consistent change in transition-only performance, and a single-fold ordinal sensitivity check does not show a clear improvement over the nominal classification approach used throughout.