Machine Learning Algorithms and Fine Art Pricing: Comparative performance to regression results for the South African market, 2009-22

Traditionally, hedonic regression and repeat sales have been used for art pricing analysis. Recently, however, machine learning algorithms have increasingly been used to provide price predictions. Using a comparative approach on a unique dataset comprised of 43,354 fine art auction lots, of which 26,486 resulted in sales, we find that a non-linear, non-parametric random forest regressor model outperforms both parametric hedonic regression models such as OLS, and convolutional neural network models. Consistent with this finding, the gradient boosting regressor approach allowing for the quantiles of the price distribution to be modeled, outperforms standard hedonic quantile regression approaches. The class of random forest regressor models and the gradient boosting quantile regressions thus appear to perform better than both neural network models and OLS models found in the literature.