Simple vs complex models in housing market forecasting : empirical evidence from Helsinki metropolitan area

Abstract

This study seeks to examine whether it is possible to gain similar forecasting performance from simple forecasting models compared to more complex specifications in housing market context. Evaluation is conducted by comparing the predictive power of five common modelling techniques out-of-sample: Autoregressive Integrated Moving Average (ARIMA), Simple Regression (SR), Multiple Regression (MR), Vector Autoregression (VAR) and Autoregressive Integrated Moving Average with a vector of explanatory variables (ARIMAX).

A set of macroeconomic variables is used with these different modelling techniques to generate ex-post (out-of-sample) forecasts for the housing market of Helsinki Metropolitan Area. The dataset employed in this study is gathered from public sources and covers a period from 1999 to 2018. The ex-post forecasts are generated one, two, three, four and five steps ahead, i.e. from 2016 H2 to 2018 H2, and the forecasting accuracy is assessed by calculating Theil’s U and root-mean-square error (RMSE) values for each of the forecasts.

The obtained results imply that added model complexity does not necessarily yield better results, as the more complex run the risk of overfitting small data samples. What is more, the results indicate that while the complex models tend to fit historic data with greater accuracy, the higher historical fit does not always translate into superior forecasting results. However, it seems probable that the shortcomings of the more complex models in this study are aggravated by the very specific features of the utilized dataset. Hence, market participants should acknowledge that the obtained forecasting results are always not only largely dependent on the chosen methodology, but also on the utilized dataset.