Multimodal Learning for Improved NFT Price Prediction

Abstract

NFTs (Non-fungible tokens) refer to digital assets in the form of art, game items and other collectibles that are encoded in smart contracts on blockchain. Starting from 2021, the NFT market has been growing exponentially. However, the overall structures, evolutions and trends of the market have not been sufficiently explored. In this study, we analyze data of 9,045 NFTs which were on sale at OpenSea in April 2021 and predict their price using both visual and non-visual information by a two-stage machine learning approach and an end-to-end deep learning approach. To examine the effectiveness of multimodality, models trained on unimodal and multimodal data are compared. Besides, the effect of different feature fusion techniques on model performance is analyzed. To identify important predictors, the top ten most influential features are also presented. Results show that among two-stage models, the VGG-RF model trained on multimodal data gives the best performance of R squared 69.75% and RMSE 565.49. Among end-to-end models, the multimodal neural network with self and cross-attention achieves the best performance of R squared 69.25 % and RMSE 569.13. Our results also demonstrate that the visual aspect of NFT have an impact on its price, but this impact is weaker than some non-visual information such as total number of bids in history and the collection the NFT comes from. We expect findings of this study facilitate future researches on NFT trading and pricing strategies, and help investors make more advisable investment decisions.