From Causes to Forecasts: Granger Causality and Machine Learning Predictions of Housing Sales in Türkiye
2602-3954
Abstract
This study aims to examine the impact of macro-financial indicators on the housing sales volume in Türkiye and evaluate the forecasting performance of traditional time-series models versus modern machine-learning algorithms. Using monthly data from January 2014 to November 2024, the research uses linear and non-linear Granger causality tests to investigate the lagged effects of 13 economic variables, including mortgage rates, loan volume, housing price indices, inflation, industrial production, consumer confidence, and unemployment. The findings reveal statistically significant causal relationships between housing sales and several indicators, notably interest rates, housing prices, consumer prices, and industrial output. These links imply that the cyclical effects of credit conditions, price expectations, or real-sector activity on housing demand operate with relatively short time lags. The predictive performance was evaluated using six models, including a multilayer perceptron, random forest, polynomial regression, gradient boosting, seasonal LSTM, and SARIMAX. Considering the test MAPE, the neural-network models provide the best predictions, with the multilayer perceptron and the seasonal LSTM obtaining MAPEs of 17.3% and 19.7%, respectively. On the other hand, the SARIMAX and tree-type models have worse generalisation capability. The results demonstrate the value of combining causal analysis with advanced forecasting to capture the dynamics of the housing market. They provide a practical framework for anticipating changes in demand and support the integration of machine-learning tools into economic monitoring and policy evaluation. This dual approach enhances the understanding of how economic conditions propagate through the housing sector and contributes to more informed housing market governance.