Last update: 14 Oct, 2026

<aside> <img src="/icons/list-indent_lightgray.svg" alt="/icons/list-indent_lightgray.svg" width="40px" />

Table of contents

</aside>

<aside> <img src="/icons/user-circle_lightgray.svg" alt="/icons/user-circle_lightgray.svg" width="40px" /> Available for everyone

</aside>

<aside> <img src="/icons/copy_lightgray.svg" alt="/icons/copy_lightgray.svg" width="40px" /> Related articles Buy an artwork

</aside>

In this article An overview of artwork valuation models: how linear regression and XGBoost are used and how these models aim to bring structure and liquidity to the art market.

What are Artwork Valuation Models?

These are our proprietary valuation models, which were inspired by asset pricing models used in financial markets. They combine logarithmic regressions, elements of the Black-Scholes option pricing model and other mathematical tools to calculate the price of an art object.

We used a dataset of global sales numbers since the year 2000 for all artworks priced below $30k. The dataset size is approximately 300,000 artworks. Across all models, and wherever applicable, R² ranges from 61% to 80%, with the XGBoost model showing the strongest potential to explain prices.

We have developed five distinct models, with the results of the two best-performing ones available for your consideration next to each artwork's price. We are working on several improvements and hope to release the forward-looking model in 2027, which will include each customer's subjective utility function.

These models were developed together with the Department of Physical and Mathematical Sciences at the National University of Kyiv-Mohyla Academy.

Note: We do not use AI in our valuation models.

Our current Selection of Models

• Simple Linear Regression: This is the main model for the econometric part of the study. Its coefficients allow for an explanation of how individual characteristics relate to changes in price. However, for practical forecasting, it is inferior to boosting models.
• XGBoost Multiple Decision Trees: The primary model for practical price forecasting. It achieves the lowest MAE and RMSE, as well as the highest R2 value. It is particularly convenient for working with a large number of categorical features, as it handles them without requiring prior one-hot encoding.

Why does the XGBoost model perform better?

The price of an art object depends not only on its individual characteristics, but also on complex interactions between them and external factors. For example:

• the impact of the material may depend on the category of the work;
• the effect of size may differ for painting, sculpture, and photography and is highly non-linear;
• the significance of nationality or style may depend on the author;

Linear Regression primarily assumes a linear and additive relationship between features and the target. XGBoost automatically discovers non-linearities and feature interactions, which explains its superior results.

Why we use valuation models?

The primary goal of these models is to bring structure to the market as a whole and improve liquidity. We will continue refining them to provide buyers with additional tools for valuing art based on personal utility functions and future market expectations.

<aside> 🖐️ Can’t find an answer to your question?

Please get in touch with us via [email protected]

</aside>