Deep Learning Boosts Valuation Accuracy for First‑Time Art Sales

Global: Deep Learning Boosts Valuation Accuracy for First‑Time Art Sales

A study posted to arXiv on December 28, 2025, by Jianping Mei, Michael Moses, Jan Waelty, and Yucheng Yang investigates how deep learning can improve price predictions in the art market. The authors compare traditional hedonic regression and tree‑based models with modern multi‑modal neural networks that combine tabular auction data and visual embeddings of artworks. Their goal is to assess whether visual information can add economic value when historical price anchors are missing.

Methodology

The researchers implement several deep architectures, including convolutional neural networks for image feature extraction and transformer‑based fusion layers that integrate those features with auction metadata such as artist name, provenance, and prior transaction history. Model performance is evaluated using out‑of‑sample mean absolute error (MAE) and R² metrics across multiple test splits.

Dataset

The analysis draws on a large repeated‑sales dataset compiled from major auction houses, covering thousands of transactions over the past two decades. For each sale, the dataset includes high‑resolution images of the artwork, sale price, and detailed descriptive attributes. A subset of the data isolates “first‑time” sales—works with no prior auction record—to test the models under conditions where historical pricing is unavailable.

Key Findings

Results indicate that while artist identity and transaction history remain the strongest predictors overall, the inclusion of visual embeddings reduces prediction error by 7.3 % for first‑time sales compared with the best traditional baseline. For repeat sales, the improvement is modest (1.2 % reduction in MAE), suggesting that visual cues are most valuable when the market lacks comparable precedents.

Interpretability

To understand what the models learn, the authors apply Grad‑CAM visualizations and examine embedding clusters. The analysis reveals that the networks attend to compositional elements such as brushstroke texture, color palette, and subject matter—features that align with expert appraisals of artistic style and quality.

Implications

According to the authors, the findings offer a data‑driven tool for auction houses, insurers, and collectors seeking more reliable valuations of novel artworks. By quantifying visual contributions, stakeholders can better assess risk and price uncertainty in markets traditionally reliant on expert judgment.

Limitations and Future Work

The study acknowledges that the dataset is limited to publicly reported auction results, potentially excluding private sales that could affect model generalizability. The authors propose extending the framework to incorporate provenance documents and textual descriptions, as well as testing the approach on other cultural‑asset markets.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.