Machine Learning-based Prediction of House Sale Prices in Hulu Langat

Abstract

House price prediction remains a complex task due to the interplay of structural, locational, and economic factors. This study proposes a machine learning-based predictive framework tailored to the residential real estate market in Hulu Langat, Malaysia, a district undergoing rapid urbanization. Using open-source datasets, three structured variations are developed: one containing only housing and locational attributes, another incorporating regional-level macroeconomic data, and a third combining both regional and national macroeconomic indicators. The proposed model (Ensemble Weighted Average) is trained and evaluated using evaluation metrics, then compared with models such as Random Forest, XGBoost, LightGBM, and Ensemble Stacking. The proposed model trained solely on housing and locational data achieved the highest accuracy, outperforming all other models across most evaluation metrics, while XGBoost achieves the fastest computation time. Models trained with the inclusion of macroeconomic indicators consistently underperforms, suggesting that macroeconomic indicators added noise to model prediction, potentially due to spatial resolution mismatches or multicollinearity. The interpretability of the best-performing model was further enhanced with SHapley Additive exPlanations (SHAP), the resulting SHAP analysis reveals that land parcel area, property type, and local housing supply are the top contributing features to model’s performance. These findings validate the effectiveness of ensemble models for localized price prediction and highlight the importance of house attributes over broader economic trends. The proposed framework yields a practical and interpretable approach to house price prediction and may assist policymakers, developers, and planners in making informed decisions.