Abstract
We study the accuracy and usefulness of automated (i.e., machine-generated) valuations for illiquid and heterogeneous real assets. We assemble a database of 1.1 million paintings auctioned between 2008 and 2015. We use a popular machine-learning technique—neural networks—to develop a pricing algorithm based on both non-visual and visual artwork characteristics. Our out-of-sample valuations predict auction prices dramatically better than valuations based on a standard hedonic pricing model. Moreover, they help explaining price levels and sale probabilities even after conditioning on auctioneers’ pre-sale estimates. Machine learning is particularly helpful for assets that are associated with high price uncertainty. It can also correct human experts’ systematic biases in expectations formation—and identify ex ante situations in which such biases are likely to arise.
Introduction: The Challenge of Valuing Real Assets
Valuing "real assets"—such as real estate or fine art—is notoriously difficult. Unlike liquid stocks, these assets are heterogeneous (every painting is unique) and illiquid (they trade infrequently). Historically, this has made "human eyes" and expert intuition seem indispensable. This research investigates whether machine learning (ML) can capture the complex, non-linear relationships inherent in art pricing and whether it can identify patterns that human experts miss.
Methodology: Four Perspectives on Value
The study compares four distinct valuation approaches using a massive proprietary dataset from the Blouin Art Sales Index (BASI), covering 1.1 million auction lots across 371 auction houses worldwide:
- Hedonic Regression ($p_{HR}$): The traditional academic standard; a linear model based on basic features like size, artist, and medium.
- Neural Networks ($p^{txt}_{ML}$): A multi-layer perceptron trained on textual and numerical metadata (artist nationality, creation year, title keywords).
- Convolutional Neural Networks ($p^{img}_{ML}$): A computer vision approach that "looks" at the painting’s image to extract visual features, combined with metadata.
- Human Experts ($p_{AH}$): The pre-sale estimates published by auction houses like Sotheby’s and Christie’s.
Crucially, the ML models were restricted to "exogenous" data—they did not have access to the specific auction house or location, information that human experts often use to anchor their estimates.
Key Findings: Man vs. Machine
Superior Predictive Power
The ML models demonstrated a dramatic improvement over traditional statistical methods. While the standard hedonic model achieved an $R^2$ of only 0.047, the ML model reached 0.741. This suggests that neural networks are far better at capturing the "long tail" of the market—such as the specific interaction between an artist's specific "period" and the artwork's characteristics that creates exponential value.
The Human Edge and Machine Complementarity
Human auctioneers still maintain the highest overall accuracy ($R^2 = 0.912$), likely due to access to "soft" information the machine lacks: the physical condition of the canvas, provenance (prestige of previous owners), and direct market sentiment. However, the machine is not merely a subset of human knowledge. It explained 4.1% of the price variation that humans missed. Integrating ML valuations with human estimates made price predictions 18.3% more likely to be accurate than using human estimates alone.
Identifying Bias and Predicting "Buy-Ins"
One of the most significant contributions of the ML model is its ability to predict "buy-ins"—instances where a painting fails to meet its secret reserve price and goes unsold. The researchers found that human experts are often "sticky," showing reluctance to adjust valuations downward even when an artist’s market is cooling. When the human estimate was high but the ML valuation was low, the probability of a "buy-in" nearly doubled. This indicates that ML can identify systematic human biases in real-time.
Conclusion
While human experts remain the gold standard for art valuation due to their access to qualitative data, machine learning provides a powerful and objective "second opinion." It is particularly effective for high-uncertainty assets where human intuition is most likely to falter. As these technologies evolve, ML valuations may serve as a crucial check against human over-optimism and a tool for increasing transparency in historically opaque markets.
Summary updated