Predicting Housing Market Sentiment: The Role of Financial, Macroeconomic and Real Estate Uncertainties

Abstract

Sentiment indicators have long been closely monitored by economic forecasters, notably to predict short-term moves in consumption and investment. Recently, housing sentiment indices have been developed to forecast housing market developments. Sentiment indices partly reflect economic determinants, but also more subjective factors, thereby adding information, particularly in periods of uncertainty, when economic relations are less stable than usual. While many studies have investigated the relevance of sentiment indicators for forecasting, few have looked at the factors which shape sentiment. In this paper, we investigate the role of different types of uncertainty in predicting housing sentiment, controlling for a wide set of economic and financial factors. We use a dynamic model averaging/selection (DMA/DMS) approach to assess the relevance of uncertainty and other factors in forecasting housing sentiment at different points in time. We find that housing sentiment forecast errors from models incorporating uncertainty measures are up to 40% lower at a two-year horizon, compared with models ignoring uncertainty. We also show, by examining DMS posterior inclusion probabilities, that uncertainty has become more relevant since the 2008 global financial crisis, especially at longer forecast horizons.

Keywords:

• Housing sentiment
• Uncertainty
• DMA
• DMS

JEL Classifications:

• C53
• E44
• R31

Acknowledgement

We would like to thank an anonymous referee for many helpful comments. However, any remaining errors are solely ours. Further, the views expressed in this paper are those of the authors and do not necessarily reflect those of the Organisation for Economic Co-operation and Development (OECD) or the governments of its member countries.

Notes

1 This is similar to the standard state-space version of the time-varying parameter model widely used in the literature (Boivin, 2006; Kishor and Marfatia, 2013; Korobilis, 2013; Marfatia, 2014, 2015; Marfatia et al. 2017, 2019, Mbarek et al., 2019).

2 One approach is to specify a Markov switching transition matrix which requires estimating transition probabilities of switching from model $i$ to model $j$ at each time period $t$ for all the $k$ model. However, with several predictors and a large transition matrix, this can become infeasible, lead to imprecise estimates/inference, and imply a significant computational burden. See, Chen and Liu (2000) and Koop and Korobilis (2012) for more details.

3 For example, for quarterly data, when the value of $\lambda$ is set to 0.99 (0.95), it implies that the observations five years ago get 81.8% (35.8%) of the weight of the last period’s observation.

4 The data is available for download from: https://www.dropbox.com/s/al3sddq1026xci2/Online%20data.xlsx?dl=0.

5 The MU and FU indices are available for download from: https://www.sydneyludvigson.com/macro-and-financial-uncertainty-indexes.

6 The REU index is available for download from: https://sites.google.com/site/johannespstrobel/.

7 The factors are available for download from: https://www.sydneyludvigson.com/data-and-appendixes, with the underlying data derived from the FRED-MD database developed by McCracken and Ng (2016). The dataset represents broad categories of macroeconomic time series. The majority of these are real activity measures: real output and income, employment and hours, real retail, manufacturing and trade sales, consumer spending, housing starts, inventories and inventory sales ratios, orders and unfilled orders, compensation and labor costs, and capacity utilization measures. The dataset also includes commodity and price indexes and a handful of bond and stock market indexes, and foreign exchange measures.

8 To maintain brevity, we do not include plots of posterior inclusion probabilities of real house price changes and 8 factors estimated from other model specifications because they are both quantitatively and qualitatively similar across model specifications. However, these are available upon request.

9 We also considered the possibility that the influence of a positive change in uncertainty may be different from that of a negative change. Hence, we estimated all models by including positive and negative change in uncertainty as separate series in the same model. We also estimated models including only one direction of change in uncertainty. Both these exercises do not necessarily provide evidence of gains to be made by modeling asymmetric effects of uncertainty, barring a few instances of weak (with the MSE-F test being significant at the 10% level) gains at $h$=1. These results are not included to maintain brevity, but are available upon request from the authors.

10 The alternative series in the literature, for example, the Soo (2018) dataset, are even shorter (starting in 2000 to 2013) than NAHB data series. Using such short data series within the Dynamic Factor Model of the paper may raise concerns over the reliability of the results and inferences therefrom. However, we still test for Granger causality, as we do for other measures, and the results from the available 56 observations are presented in Appendix C. Again, consistent with our basic story, we find that uncertainty at longer-lags tend to carry in-sample predictive information for Soo’s (2018) sentiment index.