Constructing a multi-family housing sales price index using machine learning: Focusing on small-area units in Seoul

This study highlights the importance of constructing reliable price indices for multi-family housing and empirically examines the feasibility of developing such indices at both citywide and subregional levels within Seoul. Using LightGBM, a tree-based machine learning model well-suited for capturing heterogeneous submarket dynamics, the study estimated monthly transaction prices for individual multi-family units in Seoul from 2006 to 2024, based on publicly disclosed actual transaction data and housing registry records. The final model achieved a prediction error of 9.31%, demonstrating strong performance in tracking high-resolution price levels and temporal changes across the city. Notably, SHAP analysis revealed that price increases were concentrated in areas undergoing housing redevelopment. The Seoul-wide index constructed in this study more closely resembled transaction-based indices than appraisal-survey- based indices, offering a more accurate reflection of market trends. At the subregional level, the machine learning?derived indices show considerable potential for policy applications, such as early detection of real estate fraud and informing targeted public interventions. Future research should extend this approach to detached and single-family homes and investigate spillover effects from nearby apartment prices to enhance understanding of low-rise residential markets.