Real Estate Price Prediction in Dubai Using Machine Learning Models with Stacking Ensemble Technique

Abstract

This study empirically compares and applies several machine learning models to estimate real estate prices in Dubai. The study also investigates the effect of feature engineering and temporal dynamics on models’ performance. It uses a large-scale dataset of 610,089 residential property transactions from the Dubai Land Department (DLD) / Dubai Pulse open-data repository over the period 2015–2024. Addressing a gap identified in the literature, its contribution lies in combining temporal feature engineering (lag and rolling indicators), leakage-aware time-series validation, stacking ensemble evaluation, and valuation-specific metrics in the Dubai context. We test several machine learning models including Random Forest, Gradient Boosting, LightGBM, CatBoost, Ridge Regression, K-Nearest Neighbors as well as several different versions of a stacking ensemble. The models are tested using the standard metrics for machine learning, R², MAE, RMSE, MAPE as well as two metrics specific to financial/valuation modeling, COD and PRD. The experiments show that among the considered machine learning models the tree-based ensemble models and the stacking ensemble models clearly outperform the linear and the distance-based models. The two-stage stacking ensemble attains the highest coefficient of determination (R² = 0.8109), while CatBoost achieves the lowest error (RMSE = 660,702; R² = 0.8097) and LightGBM reaches R² = 0.8025, with feature engineering based on lag and rolling features contributing substantially to these results. In valuation terms, the leading models achieve strong consistency and minimal systematic bias, with LightGBM attaining the lowest coefficient of dispersion (COD = 13.80%) and the stacking ensemble maintaining a price-related differential close to one (PRD = 1.049). The results of the feature importance show that the structural characteristics of a property (especially the size) play the biggest role for the pricing of a property. Temporal and spatial features provide additional information that, although less explanatory for individual forecasts, improves overall forecasting performance.