Abstract
Non-fungible tokens (NFTs) have emerged as a rapidly growing digital asset class, enabling new forms of ownership, exchange, and monetization of digital content. Artists can monetize their work, users can own virtual assets, and digital items can be uniquely created and traded on blockchain platforms. This study investigates the dynamics of daily NFT transaction volumes over the period 2021–2022, focusing on three market segments: collectible, game, and utility NFTs. We compare the predictive performance of two widely used time-series models, the AutoRegressive Integrated Moving Average (ARIMA) and the Heterogeneous AutoRegressive (HAR) model, estimated both with and without sentiment and attention indicators derived from textual data and online search activity. The results show that incorporating sentiment and attention indicators improves in-sample model fit for collectible and game NFTs. Evidence of enhanced out-of-sample forecasting performance is found only for the collectible sector. For game NFTs, the contribution of sentiment variables is not supported by statistically significant improvements in predictive accuracy. In contrast, utility NFTs appear to be primarily driven by their own past transaction dynamics, with no evidence that sentiment indicators improve either model fit or forecasting performance. These findings suggest that the informational content of media sentiment varies across NFT categories and is more relevant in segments where market dynamics are strongly influenced by social interaction and investor attention. The results provide useful insights for researchers, policymakers, and market participants interested in understanding and forecasting NFT transaction activity.