Recurring patterns in the run-up to house price busts

Abstract

We present evidence that shows that large increases in credit and residential investment shares, along with deteriorating current account balances, provide useful leading indicators of house price busts. These variables also explain cross-sectional patterns in the build-up to the 2007 crisis. Interestingly, movements in output and inflation have little ability to predict house price busts.

Acknowledgements

The authors thank Olivier Blanchard, Charles Collyns, Jörg Decressin, Antonio Fatás, Jordi Galí, David Romer and seminar participants at the IMF Brown-Bag Seminar for thoughtful discussions and comments. The remaining errors are our own. Gavin Asdorian and Jessie Yang provided exceptional research assistance. The views expressed in this article are those of the authors and do not necessarily represent those of the IMF or IMF policy.

Notes

¹See also Berg and Pattillo (1999).

²For our sample of countries, the average growth rate of house prices is 2.4%, with an SD of 8%.

³Subject to data limitations, the sample includes the following countries: Australia, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, the Netherlands, New Zealand, Norway, Spain, Sweden, Switzerland, the United Kingdom and the United States. Data on real house prices were obtained from the Organization for Economic Cooperation and Development (OECD).

⁴The duration of a bust is the amount of time the four-quarter moving average of the growth rate of the asset price remains below the relevant threshold. Because periods t−3 to t are labelled as a bust, there is a minimum duration of 1 year for all busts.

⁵Trend output is measured using a one-sided HP filter with a smoothing coefficient of 1600.

⁶The data source for all variables, except credit, is the OECD Analytical Database. Data on domestic credit to the private sector were obtained from the IMF's International Financial Statistics.

⁷The noise-to-signal ratio is typically defined as [B/(B + D)/A/(A + C)] (see Table 2 for the classifications). To avoid the influence of extreme observations, we limit our grid search to four percentiles: 70th, 75th, 80th and 90th. The percentiles for each indicator are computed based on a ‘real time’ approach, using observations over the previous 15 years (see Alessi and Detken, 2009). As such, the statistics are calculated only for the post-1985 period.

⁸In terms of the matrix presented in Table 2, this statistic can be computed as A divided by (A + B).

⁹In this case, the relevant statistic is C divided by (A + C).

¹⁰In the sample, the unconditional probability of a house price bust occurring 1–3 years in the future is 14% during the post-1985 period.

¹¹Probit models have been used in the context of predicting currency crises (Frankel and Rose, 1996; Milesi-Ferretti and Razin, 1998).

¹²Variables are measured as deviations relative to the rolling HP filter, as used earlier.

¹³The run-up period was chosen to be from 2002 to 2006, as it represented the period over which there was a uniform house price appreciation across all countries (with the exception of Germany and Japan, which have experienced secular declines).