A Bayesian dynamic hedonic regression model for art prices

Abstract

Traditional ordinary least squares (OLS) regressions applied to hedonic pricing models assume that, when using time series, the estimated coefficients for each of the attributes remain constant. We propose a Bayesian dynamic estimation of the hedonic regression model in which the estimated coefficients can be time-varying, demonstrated with an application of art prices. Our dynamic linear regression model overcomes the problems associated with traditional rolling-window based OLS (which represent ad hoc approximations to dynamic estimation), such as under or over-estimation of parameter values and non-adaptive window sizes to account for time-variability. Using a sample of 27,124 paintings sold at auction from 63 Pop-artists (2001–2013), we demonstrate that the estimated coefficients associated with commonly used art attributes fluctuate noticeably through time, and that certain types of artworks and artists might be regarded as “safer” investments (as their art experiences smaller maximum drawdowns), based on price dynamics during the financial crisis (2008–09).