Predicting Commercial Performance from Land Use Configuration: A Spatially and Temporally Validated Machine Learning Framework for Seoul’s Commercial District Census

Predictive models built from administrative urban data are increasingly used to guide planning decisions, but their reported accuracy is trustworthy only if it is evaluated on places and years the model has not seen. We develop such a framework for a concrete question: how commercial performance varies with the intensity and extent of urban commercial land. Using two prediction waves from Seoul’s commercial district census (2023→2024 and 2024→2025, 3044 district waves from 1530 districts), an XGBoost model interpreted with SHAP and accumulated local effects attains R2 = 0.325 under nested leave-one-gu-out spatial cross-validation, 0.532 in temporal transfer, and 0.299 in a fully nested combined spatio-temporal holdout, while random cross-validation inflates the spatial figure to 0.517. An untransformed linear model collapses to 0.007, but spline regression on transformed predictors reaches 0.354; the gain comes from nonlinearity, not boosting. Sales per store rise with store density up to an exhaustion threshold near 44 stores/ha and are flat beyond it; the district area profile begins to flatten around 14 ha. The centred ALE curves cross zero near 25 stores/ha and 6 ha, which are sample-relative reference points, not economic thresholds. Alley districts sit mostly below the density reference point, traditional markets are distinguished by small land extent rather than intensity, and developed districts are locally unpredictable. Within-district first differences do not reproduce the cross-sectional density gradient, so these quantities are configurational benchmarks, not causal effects or intervention targets.