Forecasting US housing starts under asymmetric loss

Abstract

Survey data of forecasts of the housing market may provide a particularly rich data environment for researchers and policymakers to study developments in housing markets. Based on the approach advanced by Elliott et al. (2005), we studied the properties of a large set of survey data of housing starts in the United States. We document the heterogeneity of forecasts, analyse the shape of forecasters' loss function, study the rationality of forecasts and the temporal variation in forecasts.

Keywords:

• housing starts
• loss function
• rationality of forecasts

JEL Classification::

• D84

Acknowledgements

We thank the Fritz-Thyssen-Stiftung for financial support (AZ.10.11.1.167). The usual disclaimer applies.

Notes

¹ See Cain and Janssen (1995) for an early study of predicting real estate prices under the asymmetric loss. See Skitmore and Cheung (2007) for an analysis of asymmetric loss functions for construction-price forecasting.

² Elliott et al. (2005) suggest to use the lagged forecast error as an additional instrument. We do not use the lagged forecast error as an instrument because our survey data consists of an unbalanced panel of forecasts of housing starts.

³ All figures and all computations were implemented using the software R (R Development Core Team, 2012).

⁴ Because of the boom-bust cycle, we assume that housing starts can be treated as being stationary during the sample period that we study. The alternative would be to perform our analysis in terms of growth rates. Forecasters, however, forecast the level, not the growth rate of housing starts.

⁵ The kind of cross-sectional scattering of forecasts as illustrated in Fig. 1 has also been analysed in earlier empirical studies of, for example, exchange-rate forecasts (e.g. Benassy-Quere et al., 2003) and forecasts of inflation rates (e.g. Capistrán and Ramos-Francia, 2010). Capistrán and Timmermann (2009) study the link between the cross-sectional range of forecasts and the asymmetry of forecasters' loss functions in their study of inflation forecasts.

⁶ As expected, the estimated asymmetry parameter looked more stable when we used a recursive-estimation window (not reported, but available from the authors upon request). We also observed, however, a tendency of the asymmetry parameter to increase (albeit more moderate than in Fig. 4) at the end of the sample period when we used a recursive-estimation window.