基于机器学习的土地估值方法

With the rapid increase of transaction volume in the land market and frequent usage of land as financial assets, traditional methodologies for determining land value are gradually unable to meet the needs of the current market due to their high cost and strong subjectivity. Therefore, cheap and efficient intelligent approaches are urgently needed. Based on previous literature, this paper takes land’s basic characteristics, business information nearby, and local macroeconomic information as the main pricing factors, and applies Linear Regression, Decision Tree, Artificial Neural Networks, and Deep Learning model to land evaluation. The data for training testing the models is composed of 58, 815 parcels of land traded through bidding, auction, and listing, collected from Chinese websites in the period of January 2016 to June 2019. The study demonstrates that among those models, XGBoost outperforms all the other models and is well suited to different types of land. Furthermore, we measure the influence of enriched attributes on model performance and find that all three types of factors are indispensable for determining the price of land, especially the max floor area ratio and cities land prices, which means the macro-market value of the land and the degree of land availability are the most important factors.