Combination of machine learning-based automatic valuation models for residential properties in South Korea

Abstract

The applicability of machine learning (ML) techniques has recently been expanding to include automatic real estate valuation models. The main advantage of this technique is that it can better capture complexity in the value determination process. Therefore, the performance of these techniques is shown to be superior to conventional models. In this paper, the latest ML algorithms (i.e., support vector machine, random forest, XGBoost, LightGBM, and CatBoost algorithms) are examined as automatic valuation models, and several combination methods are proposed to improve the models’ predictive power. We applied ML models to approximately 57,000 records on apartment transactions, which were provided by South Korea’s Ministry of Land, Infrastructure, and Transport, that occurred in Seoul in 2018. The results are as follows. First, ML-based predictors (especially, the latest decision tree-based algorithms) are more performative than conventional models. Second, the prediction error from a model can be partially offset by another model’s error, which implies that an efficient averaging of the predictors improves their predictive accuracy. Third, the models’ relative performance may be relearned by the ML algorithms, which means that they can also be used to recommend which algorithm should be selected for making predictions.

Keywords

• automatic valuation model
• mass appraisal
• machine learning (ML) techniques
• combined approach
• decision tree-based algorithms