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

Abstract Images and categorical traits provide closely related descriptions of profile picture non-fungible tokens (NFTs), making the incremental predictive value of image embeddings uncertain. We compare metadata-only, image-only, early-fusion and weighted late-fusion models for Bored Ape Yacht Club (BAYC) and Mutant Ape Yacht Club (MAYC) using eight regression families and seven visual encoders. Models are selected on pre-2025 data and retrospectively evaluated from January 2025 to April 2026. Selected models in all four groups outperform a zero log-relative-price benchmark. Relative to metadata, early fusion reduces root mean squared error (RMSE) by 0.09% in BAYC and 0.95% in MAYC; late fusion increases BAYC RMSE by 0.07% and reduces MAYC RMSE by 0.70%. Sensitivity analyses assess unseen tokens, matched regression families, transaction-row versus count-weighted fitting, and full reselection under stricter cleaning. For unseen tokens, exploratory 95% fusion-versus-metadata intervals include zero in both collections. Metadata-only partial least squares (PLS) outperforms CLIP–PLS early fusion in MAYC. Across two stricter cleaning variants, 13 of 16 representative configurations change relative to a common base-clean control; fusion improvements are not consistently supported after reselection. Some tree-based comparisons are implementation-sensitive. The results do not establish consistent image gains, independent aesthetic pricing power, or superiority of either fusion strategy.