Modern art meets machine learning: advancing price predictions with visual features

Abstract

This thesis explores machine learning's (ML) potential in art price prediction by integrating traditional artwork features with visual data from high-quality images. Using a dataset of modernist artworks auctioned by Christie’s (2007-2024), the study employs hedonic regression, advanced ML, and Convolutional Neural Networks. Results show ML models outperform traditional methods, revealing biases in auction house estimates. Visual features are relevant predictors of price and improve predictions modestly. Auction house valuation biases when combined with ML modes can be interpreted and subsequently mitigated. These findings highlight the promise of data-driven, scalable art valuation frameworks, advancing both academic research and industry practice.