Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Deep Learning for Art Market Valuation

Jianping Mei, Michael Moses, Jan Waelty, Yucheng Yang

TL;DR

This paper investigates whether visual content from artworks, processed through a multi-modal deep learning framework, adds predictive value to art-market valuations beyond structured predictors and prior transactions. It compares hedonic regressions, XGBoost, and two-branch neural nets that fuse image embeddings with metadata, under a forward-looking temporal split that separates fresh-to-market from repeated sales. The key finding is state dependence: price history and reputation dominate valuations for repeat sales, while image-derived signals meaningfully improve predictions for first-time sales, with embedding dimension playing a bias-variance role; interpretability analyses (Grad-CAM and PCA embeddings) reveal that visual cues relate to composition and texture and interact with object-level features. Although presale estimates remain the strongest predictor, combining them with ML signals improves calibration and reduces tail biases, offering practical guidance for valuation practice and advancing understanding of when unstructured visual data can aid asset pricing. The methodology, findings, and interpretability analyses contribute to the literature on asset pricing with unstructured data and provide actionable insights for art-market practitioners.

Abstract

We study how deep learning can improve valuation in the art market by incorporating the visual content of artworks into predictive models. Using a large repeated-sales dataset from major auction houses, we benchmark classical hedonic regressions and tree-based methods against modern deep architectures, including multi-modal models that fuse tabular and image data. We find that while artist identity and prior transaction history dominate overall predictive power, visual embeddings provide a distinct and economically meaningful contribution for fresh-to-market works where historical anchors are absent. Interpretability analyses using Grad-CAM and embedding visualizations show that models attend to compositional and stylistic cues. Our findings demonstrate that multi-modal deep learning delivers significant value precisely when valuation is hardest, namely first-time sales, and thus offers new insights for both academic research and practice in art market valuation.

Deep Learning for Art Market Valuation

TL;DR

This paper investigates whether visual content from artworks, processed through a multi-modal deep learning framework, adds predictive value to art-market valuations beyond structured predictors and prior transactions. It compares hedonic regressions, XGBoost, and two-branch neural nets that fuse image embeddings with metadata, under a forward-looking temporal split that separates fresh-to-market from repeated sales. The key finding is state dependence: price history and reputation dominate valuations for repeat sales, while image-derived signals meaningfully improve predictions for first-time sales, with embedding dimension playing a bias-variance role; interpretability analyses (Grad-CAM and PCA embeddings) reveal that visual cues relate to composition and texture and interact with object-level features. Although presale estimates remain the strongest predictor, combining them with ML signals improves calibration and reduces tail biases, offering practical guidance for valuation practice and advancing understanding of when unstructured visual data can aid asset pricing. The methodology, findings, and interpretability analyses contribute to the literature on asset pricing with unstructured data and provide actionable insights for art-market practitioners.

Abstract

We study how deep learning can improve valuation in the art market by incorporating the visual content of artworks into predictive models. Using a large repeated-sales dataset from major auction houses, we benchmark classical hedonic regressions and tree-based methods against modern deep architectures, including multi-modal models that fuse tabular and image data. We find that while artist identity and prior transaction history dominate overall predictive power, visual embeddings provide a distinct and economically meaningful contribution for fresh-to-market works where historical anchors are absent. Interpretability analyses using Grad-CAM and embedding visualizations show that models attend to compositional and stylistic cues. Our findings demonstrate that multi-modal deep learning delivers significant value precisely when valuation is hardest, namely first-time sales, and thus offers new insights for both academic research and practice in art market valuation.
Paper Structure (41 sections, 5 equations, 22 figures, 7 tables)

This paper contains 41 sections, 5 equations, 22 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Literature
  3. Data and Methodology
  4. Data
  5. Variables.
  6. Fresh-to-market versus repeated sales.
  7. Data filtering and preprocessing.
  8. Models
  9. Neural Networks
  10. Linear Regression
  11. XGBoost
  12. Estimation Procedure
  13. Train--test split and objective.
  14. Optimization and tuning.
  15. Results
  16. ...and 26 more sections

Figures (22)

  • Figure 1: Distribution of selected categorical variables
  • Figure 2: Number of transactions per transaction type
  • Figure 3: Neural network architecture with structured metadata and image features
  • Figure 4: By medium
  • Figure 5: By category
  • ...and 17 more figures