A Comparative Study of Time Series Models for Real Estate Market Forecasting

The real estate market is dynamic and influenced by various economic factors, making price forecasting a complex yet crucial task for both investors and buyers. This study proposes a robust forecasting model aimed at predicting real estate prices. Using a large dataset of over 2.2 million records from 1990 to 2023, this study employs multiple time series models including ARIMA, SARIMAX, Prophet, Holt-Winters, LSTM and a hybrid model (SARIMAX + ARIMA) to compare their effectiveness in forecasting property prices. The dataset undergoes extensive preprocessing techniques, such as resampling, scaling and stationarity checks using the Dickey-Fuller test, followed by first-order differential to address non-stationarity. The models are evaluated using performance metrics such as MAPE, MAE, MSE and RMSE to ensure comprehensive accuracy assessment. Additionally, nature-inspired optimization techniques such as Genetic Algorithms are employed for parameter tuning further enhancing model performance. Visualizations like ACF/PACF plots and residual analyses are presented to understand model fit, along with detailed predictions and trends for the next 12 years. The results demonstrate that while complex models like SARIMAX and LSTM capture long-term dependencies and seasonality, simpler models like ARIMA also perform competitively. This study concludes by recommending the best-suited model based on forecasting accuracy, thereby providing valuable insights for real estate market stakeholders.