Biased Auctioneers

Abstract

We study biases in art auction house estimates by generating our own counterfactual pre-sale valuations through a purpose-built machine-learning algorithm that considers both visual and non-visual artwork characteristics. We find that auction house estimates are not informationally efficient, as they can be improved upon as predictors of transaction prices. Moreover, estimates show systematic over- and undervaluation patterns that are predictable ex ante. Finally, we find that biases in pre-sale estimates drive market participants’ investment outcomes: higher relative estimates are associated with both lower sale rates and lower future capital gains.

Keywords: art, auctions, experts, asset valuation, biases, machine learning, computer vision

JEL Classification: C50, D44, G12, Z11