FORECASTING NEW ZEALAND HOUSING DEMAND: COMPARING ELASTIC NET, XGBOOST AND RNN MODELS

Abstract

Forecasting housing demand has been a prevalent research focus globally, primarily employing traditional econometric methods. However, the application of machine learning in this domain remains limited, particularly in the New Zealand context. This study addresses this gap by implementing Elastic Net, XGBoost, and Recurrent Neural Network models to predict residential housing demand in New Zealand using historical economic and demographic data from 1995. The models were evaluated using a comprehensive framework of six complementary metrics (R², SMAPE, MAE, RMSE, MBE, and MASE), with the RNN model achieving the highest accuracy. Results demonstrate that machine learning algorithms significantly enhance housing demand forecasting, with temporal models outperforming traditional approaches. The analysis of feature importance identified CPI, construction investment, import values, and unemployment as key drivers, while demographic factors showed limited impact on housing demand. These findings provide valuable insights for policymakers and construction firms addressing New Zealand's housing challenges. Future research should expand dataset dimensions and improve model interpretability