Given the uniqueness of fine art objects and uncertainties in demand (at auctions), can investors in paintings get accurate estimates of market values of holdings and potential acquisitions? In their March 2019 paper entitled “Machines and Masterpieces: Predicting Prices in the Art Auction Market”, Mathieu Aubry, Roman Kräussl, Gustavo Manso and Christophe Spaenjers compares accuracies of value estimates for paintings based on: (1) a linear hedonic regression (factor model), (2) neural network software and (3) auction houses. For the first two, they employ 985,188 auctions of paintings during 2008–2014 for in-sample training and 104,404 auctions of paintings during the first half of 2015 for out-of-sample testing. Neural network software inputs include information about artists and paintings (year of creation, materials, size, title and markings), and images of the paintings. Using information about artists/paintings and images and auction house estimates and sales prices for the specified 1,089,592 paintings by about 125,000 artists offered through 372 auction houses during January 2008 through June 2015, they find that:
- In training data, average (median) sell price is $61,492 ($3,526), with a long right tail of extremely expensive paintings. About a third of paintings do not sell due to failure to reach reserve price.
- Hedonic regression indicates that paintings that are bigger, signed or dated, self-portraits and done in oils command significantly higher prices. In-sample, hedonic regression explains 79% of variation in prices among works (mostly due to the artist). But, out-of-sample the regression explains only 5% of price variation.
- Neural network software is much better at handling sale price non-linearity than hedonic regression. Neural network software using non-image information explains 73% of variation in out-of-sample prices. Adding images increases predictive power to 75%. In other words, the predictive value of images is small.
- Auctioneer estimates explain 88% of variation in out-of-sample prices. These experts have extraordinary access to information about painting provenance and quality. Also, it is possible that auction house estimates guide bidding.
- Neural network software estimates retain some predictive power after controlling for auctioneer estimates. For example:
- When the auctioneer estimate is high (low) compared to the software estimate, the likelihood of reaching the reserve price is relatively lower (higher).
- The software is particularly helpful for works of artists with very high price volatility or recent price declines unrecognized by auctioneers.
In summary, evidence indicates that neural network software is much superior to linear factor modeling and nearly as good as expert auctioneers in valuing fine paintings.
Cautions regarding findings include:
- Valuations are gross, not net. Fees for auctioning or otherwise brokering art sales may be substantial.
- The neural network software valuation approach is not available to most art investors, who would bear fees for such a valuation service.
See “Aesthetic Investments” for summaries of research on the return on art and other collectibles.
