Automating Art Appraisals: Machine Learning for Art Valuation

Abstract

A banana stuck to the wall with duct tape was sold for $150,000 in 2019 at the Art Basel Miami art fair. Is the price of artwork idiosyncratic, depending on the whims of art collectors? The booming $50 billion-a-year art market is dominated by the three largest auction houses, Sotheby’s, Christie’s, and Phillips. Each relies on the judgments of their experienced art auctioneers to appraise artwork. This thesis attempts to understand the drivers of art prices to facilitate the further growth of art markets. It seeks to answer the question: can a machine learning model mimic or even surpass the judgments of seasoned art auctioneers? We collect and clean historic auction data from Sotheby’s, Christie’s, and Phillips using the databases Artnet and MutualArt. Standard hedonic pricing models and neural networks are then constructed to model the auction sale prices on the textual, numerical, and visual features of artworks. The neural network model that combines numerical and textual and image data is found to perform well (with R-squared = 0.6180). Larger, signed paintings on canvas tend to sell at higher price points, while artworks with common titles (e.g. untitled, figure, landscape) sell at lower. The study shows that the current neural network model is not yet able to replace tedious and expensive human appraisal processes since the estimated price ranges of appraisers have an R-squared of 0.8930 and beat the current neural network model. However, we see an improvement in the predictive power of the current model compared to previous research. Our findings reinforce the challenge of integrating machine learning into the art appraisal process. The lack of accessible auction data and immense numbers of potential drivers for art prices are significant barriers.