Based on a spatial clustering and partitioned stacking ensemble model, this study addresses the limitations of traditional geoweighting regression in capturing nonlinear location premiums and submarket heterogeneity within urban real estate markets. It proposes a two - stage modeling framework: “spatial clustering → partitioned differentiated stacking ensemble.” Using long - term multi - source transaction data for Beijing's secondary housing market, the study divides the market into three spatially heterogeneous submarkets: core, near - suburban, and far - suburban. Stacked ensemble models based on ElasticNet, XGBoost, LightGBM, and Random Forest are constructed within each submarket.Factor analysis extracts interpretable common factors, which are combined with Lasso and SHAP for feature selection and impact mechanism analysis.Results indicate that the zoned stacking model performs exceptionally well across all three submarkets, achieving an R² of 0.916 in the core urban area. Significant nonlinear location premiums exist within the core urban area.The multi - level interpretability framework reveals the differentiated effects of location and scale factors across different submarkets.This study advances from “global modeling” to “spatial zoning + adaptive ensemble,” providing a viable tool for refined valuation and risk management in highly heterogeneous markets.