Forecasting Housing Prices with Network Information

Abstract

While the existing literature highlights interconnections across housing markets, it largely overlooks their potential to enhance house price predictions. To address this gap, we integrate housing market links into a large-dimensional vector autoregressive (LDVAR) model and employ shrinkage techniques for estimation. Simulation outcomes reveal that the LDVAR model's predictive efficacy hinges on network sparsity and connectivity strength. While univariate models suffice in scenarios of sparse and weak connections, incorporating network information via the LDVAR specification is crucial when the connections between housing markets are strong. Network information holds greater importance for shorter-term forecasts within denser networks, while its effect diminishes as forecasting horizons extend and networks become sparser. We examine three forecasting scenarios, each with its own degree of network connectivity, by considering house prices within Sydney and Melbourne in Australia, as well as city-level house prices in China. Our empirical findings corroborate the results of our simulation analysis.

Keywords: Housing prices, Connections, Sparsity, Large dimension, Shrinkage, Out-of-sample forecasting, Model confidence set

JEL Classification: R39, C32, C51, C53