Reliability Assessment of Machine-Learning Automated Valuation Models : Apartment-Complex Holdout Validation in the Seoul Metropolitan Area

Abstract

This study evaluates the reliability of machine-learning automated valuation models (AVMs) when repeated apartment-complex observations are explicitly controlled. AI-based AVMs refer here to random forest, LightGBM, and XGBoost, which are compared with an unpenalized hedonic OLS benchmark using 129,237 apartment transactions assigned to 2,371 composite complex groups. The source file combines 2021-2023 transactions with 2023 management-fee records, but transaction dates are unavailable; the temporal validity of some management-fee inputs therefore cannot be verified, and related results are interpreted as retrospective and conditional. Row-wise random splitting and apartment-complex holdout testing are treated as separate operational evaluation regimes with different test samples, not as a paired estimate of the pure effect of the split rule. Under row-wise random splitting, random forest yields the lowest log-price MAE (0.1026). Under complex holdout testing, OLS has the lowest point estimate (0.1839), but the difference from LightGBM (0.1960) is not statistically significant (Delta=0.0121, 95% CI [-0.0036, 0.0282], p=0.135); XGBoost and random forest have significantly larger errors than OLS. OLS calibration indicators are closer to their ideal values, although no direct between-model calibration test is conducted. Excluding management-fee variables reverses the OLS-LightGBM point ranking; this auxiliary scenario does not establish general LightGBM superiority, but shows that relative performance is conditional on temporally valid feature availability. SHAP is used only as an interpretive diagnostic, not as evidence of predictive superiority. The findings are conditional on the specified data, validation regimes, and limited hyperparameter search space.