Key Findings
Machine Learning Price Predictions
Neural network algorithms can explain approximately 74.2% of art auction price variation, outperforming traditional hedonic models (67.7%) particularly for high-volume and high-dispersion artists
Auction House Estimate Biases
Auction house presale estimates show systematic biases, with prediction errors being persistent both at artist and auction house levels, leading to predictable patterns in price-to-estimate ratios
Economic Impact
Higher automated valuations relative to auction house estimates are associated with lower buy-in rates (25% vs 45%) and better post-acquisition returns
Predictive Power of Machine Learning vs Hedonic Models
- Machine learning algorithm explains 74.2% of price variation in test data
- Traditional hedonic model explains 67.7% of price variation
- Difference particularly pronounced for artists with more sales and price dispersion
Buy-in Rates by Machine Learning Valuation
- 45% buy-in rate when ML valuation is low relative to estimate
- Only 25% buy-in rate when ML valuation is high relative to estimate
- Demonstrates predictive power of ML valuations for auction outcomes
Prediction Error Variables Impact
- Benchmark model explains 6.7% of prediction error variation
- Removing auction information reduces explanatory power to 3.4%
- Model without artist information explains only 1.2% of variation
Contribution and Implications
- First study to systematically examine the predictability of art auction house prediction errors
- Demonstrates potential for machine learning to complement human expertise in art valuation
- Provides evidence of systematic biases in auction house estimates that affect market outcomes
- Shows practical applications for automated valuation methods in illiquid asset markets
Data Sources
- Model comparison chart based on R-squared values from Table III and regression results
- Buy-in rates visualization derived from Figure 10 data showing relationship between ML valuations and auction outcomes
- Prediction error variable impact chart constructed using R-squared values from Table VI