Text Mining Based Price Prediction in Public Auctions using Land Appraisal Reports

This study aimed to predict land prices of public auctions based on appraisal reports. Analyzing 24,047 land transactions from Onbid (2017‒2024), the dataset included property details, bidding information, appraisal reports and regional data on population·households (2,784 cases), local taxes (2,080) and land prices (28,235). Text mining extracted qualitative variables, while instrumental variable and two-stage least squares investigated the causation of auction prices. Performances were assessed by adj.R², root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), successful bids and probability of successful bids. Predictions for the rounds of bidding and the prices of land categories indicates that MAPE performed better in quantitative analysis, whereas MAE and RMSE were superior in integrated analysis. The probability of successful bids was about 2.89% to 22.15% higher for integrated analysis cases on every case. Quantitative variables (e.g.,the ratio of the prices on public auctions in the neighborhood) were effective with ample historical data and stable economic conditions. In contrast, qualitative factors (e.g., transportation, sentiment and legal issues) were valuable either data were limited or economic conditions were unstable. This study pioneered how to determine qualitative variables by analyzing appraisal reports and identified the characteristics of quantitative and qualitative factors to propose optimal situation-specific predictions. Additionally, this study contributes to public auctions and general valuation of lands by deriving price factors with high explanation but low prediction errors.