Predicting Hammer Prices in Judicial Real-Estate Auctions on UYAP: A Leakage-Safe CatBoost Baseline with Robust Targeting

We present a leakage-safe, production-ready framework for predicting hammer prices in judicial real-estate auctions on Turkey's UYAP e-Sales platform. We target the log difference DIFF = log(1 + bid) - log(1 + est_val), which reduces heteroskedasticity and enables cross-period comparison. Features use only information available before auction close; macro deflation (CPI/HPI) and a past-only spatio-temporal neighborhood signal (ST-lag) are central. We train CatBoost with chronological splits (train ≤ 2023-12-31; validation 2024; test ≥ 2025). On test data, the Apartment model achieves R²≈0.70, MAE ≈1.20M TL, and MAPE ≈25%. A coarse price_band has limited global effect for Apartments, lowers MAPE but raises variance for Shops, and markedly improves RMSE and R² for Land. SHAP highlights value per m², deflated value indicators, location, deed area, and ST-lag as key drivers; Moran's I indicates no significant spatial residual dependence.