Image embeddings add little to NFT price prediction in temporal test

Profile picture NFTs such as the Bored Ape Yacht Club and Mutant Ape Yacht Club collections are described in two closely related ways: the image itself and the categorical traits recorded in their metadata. Because those descriptions overlap so heavily, researchers have debated whether visual features actually add predictive power to price models. A new study in Scientific Reports set out to answer that question with unusually careful validation.

The team, led by Geun-Cheol Lee, Hoon-Young Koo and Heejung Lee, compared four modelling strategies: metadata-only, image-only, early fusion, and weighted late fusion. They tested eight regression families and seven visual encoders, including CLIP-based representations. Crucially, the evaluation respected time.

Models were selected using data from before 2025 and then retrospectively evaluated on transactions from January 2025 to April 2026, a temporal validation scheme designed to avoid the leakage that plagues many machine-learning studies of financial markets. The headline result is modest. All four model groups beat a zero log-relative-price benchmark, but the gains from adding images were tiny.

Early fusion reduced root mean squared error by only 0.09 percent for BAYC and 0.95 percent for MAYC relative to metadata alone. Late fusion actually increased BAYC error by 0.07 percent while reducing MAYC error by 0.70 percent. Sensitivity analyses probed how robust these differences were.

The authors examined unseen tokens, matched regression families across comparisons, and alternative weighting schemes for fitting transactions. For unseen tokens, exploratory 95 percent intervals for the fusion-versus-metadata difference included zero in both collections, meaning no reliable image advantage could be established. Some comparisons proved sensitive to implementation details.

In MAYC, a metadata-only partial least squares model outperformed CLIP-based early fusion, and several tree-based comparisons changed depending on the software configuration used. The most striking finding came from stricter data cleaning. When the authors reselected models under two stricter cleaning variants, 13 of 16 representative configurations changed relative to a common base-clean control, and the apparent fusion improvements were no longer consistently supported after reselection.

The authors are explicit about the limits of their work. The results do not establish consistent image gains, independent aesthetic pricing power, or the superiority of either fusion strategy. Because the study covers two collections and a specific 2025 to 2026 evaluation window, the findings should not be generalised to the broader NFT market or to other asset classes.

Subject of Research: Predicting NFT prices from image embeddings and metadata with temporal validation

Article Title: Evaluating image embeddings and metadata for NFT price prediction using temporal validation

Article References: Lee, G.-C., Koo, H.-Y., & Lee, H. (2026). Evaluating image embeddings and metadata for NFT price prediction using temporal validation. Scientific Reports. https://doi.org/10.1038/s41598-026-77082-8

Image Credits: AI Generated

DOI: 10.1038/s41598-026-77082-8

Keywords: NFT, price prediction, machine learning, image embeddings, CLIP, temporal validation, Bored Ape Yacht Club, Mutant Ape Yacht Club, data fusion, regression models, Scientific Reports, Evaluating