We assess the accuracy and usefulness of machine-learning valuations in illiquid real asset markets. We apply neural networks to data on one million painting auctions to price artworks using non-visual and visual characteristics. Our out-of-sample automated valuations predict auction prices dramatically better than standard hedonic regressions. The discrepancies with pre-sale estimates provided by auction house experts correlate with sale outcomes: the more aggressive the auctioneer's pre-sale estimate relative to our valuation, the higher the probability of an unsuccessful auction and the lower the post-acquisition return. Finally, machine learning can detect predictability in auctioneers' "prediction errors".