Title: Statistical prediction of art prices at auction
Authors: Douglas Hodgson - UQAM (Canada) [presenting]
John Galbraith - McGill University (Canada)
Abstract: Predicting the price of a unique object, given a limited sample, is a challenging problem but of interest to market participants, including owners and insurers. The usual approach is least-squares estimation of a hedonic model for objects of a given class, such as paintings from a particular school or period. The present paper examines statistical refinements of the standard methods. First, we consider the level of aggregation that is appropriate for pooling observations into a sample, including the use of pattern recognition algorithms to identify clusters. Second, we apply model-averaging methods to estimate predictive models at the individual-artist level, in contexts where sample sizes would otherwise be insufficient to do so; averaging (ensemble prediction) is also used with regression-tree machine learning methods. Finally, we consider an additional stage in which we incorporate repeat-sale information in the subset of cases for which this information is available. The results are applied to a data set of auction prices for Canadian paintings. We compare the out-of-sample predictive accuracy of the various methods and find that those that allow us to use single-artist samples produce superior results, and that data driven averaging across predictive models produces clear gains. As well, where available,repeat-sale information appears to produce further improvements in predictive accuracy.