Reducing property valuation bias through Random Forests: Predicting prices for public asset optimization

Abstract

This study applies supervised machine learning, specifically the Random Forest regression algorithm, to predict office rental prices in DKI Jakarta. A dataset was compiled via web scraping of online property listings, incorporating features such as location, office area, number of floors, lifts, parking capacity, and building grade. Data preprocessing involved handling missing values, removing outliers, applying one-hot encoding, and normalizing the data to ensure consistency. The model was developed using the CRISP-DM framework and evaluated through an 80:20 train-test split and 10-fold cross-validation. Performance metrics included Root Mean Squared Error (RMSE) and R². The Random Forest model achieved high accuracy, with cross-validation yielding an R² of 0.934 and an RMSE of Rp16.288 per m²/month. SHAP analysis revealed that lifts, floors, parking, office area, and building grade significantly influenced predictions. Bias analysis indicated a tendency to underestimate rents for grade B and C buildings. The model was also simulated to estimate rental values of underutilized government-owned offices, supporting asset optimization amid the planned capital relocation. These results demonstrate the potential of machine learning to improve valuation practices, reduce bias, and enhance decision-making in public asset management.