XGBoost 기반 부동산 자동가치산정모형(Automated Valuation Model)의 실증 분석- 은평구 다세대주택 실거래가를 중심으로 -

This study empirically examines the predictive performance of an automated valuation model (AVM) with the XGBoost regression algorithm based on actual transaction data on multi-family housing in Eunpyeong-gu, Seoul. Unlike prior studies that typically apply Min-Max normalization to address variable-scale discrepancies, this study maintains the original transaction price units without normalization, thereby enhancing the interpretability and practical utility of the results in real-world applications. The dataset consists of 1,839 real transactions recorded between November 2023 and October 2024, of which 1,272 were used for model training and 567 for testing. The predictive model integrates a wide range of variables including physical attributes, locational and environmental factors, as well as market and macroeconomic indicators. The empirical results demonstrated a test score (normalized root mean squared error) of 0.136511, coefficient of determination (R²) of 0.747, and mean absolute percentage error of 13.6%. These results suggest that the model can approximate the actual market prices with an average accuracy of approximately 86.4%. Given that the average transaction price during the test period was approximately KRW 291.63 million, the model’s mean prediction error corresponds to approximately KRW 39.66 million. The results confirm that the XGBoost model effectively captures the nonlinear and heterogeneous nature of real estate transactiondata. Generating predictions in raw monetary units rather than in normalized values provides a more intuitive and practically interpretable AVM structure and offers a viable alternative to conventional normalization-based frameworks. Furthermore, this study empirically demonstrates the applicability and explanatory power of machine learning models in real estate valuation, contributing both academically and practically to the development of more usable, sophisticated AVM systems.