Source-linked AI summary
TweetBoost: Influence of Social Media on NFT Valuation
Arnav Kapoor, Dipanwita Guhathakurta, Mehul Mathur, Rupanshu Yadav, Manish Gupta, Ponnurangam Kumaraguru
TL;DR
NFT valuation is challenging because sales are sparse and prices volatile. The paper links Twitter and OpenSea data to study social-media signals and predict asset value. Its best Twitter–OpenSea feature ensemble reaches 69.5% accuracy in five-class classification.
Problem
Method
The paper links tweets announcing NFTs to OpenSea listings and evaluates asset-value models using Twitter, OpenSea, and NFT-image features.
Source: §1 · p. 2Results
The best ensemble of Twitter and OpenSea features achieves 69.5% accuracy in five-class asset-value classification, while image features have limited predictive power.
Source: §1 · p. 3Takeaways & Limitations
The results indicate that Twitter and OpenSea features influence model output, with branding and metadata having a stronger effect on value than NFT images.
Source: §1 · p. 3The current system would not handle outlier transactions and schemes such as a same-buyer-and-seller sale used to artificially inflate prices and gain social-media virality.
Source: §8 · p. 8
Abstract
from arXiv · showhide
NFT or Non-Fungible Token is a token that certifies a digital asset to be unique. A wide range of assets including, digital art, music, tweets, memes, are being sold as NFTs. NFT-related content has been widely shared on social media sites such as Twitter. We aim to understand the dominant factors that influence NFT asset valuation. Towards this objective, we create a first-of-its-kind dataset linking Twitter and OpenSea (the largest NFT marketplace) to capture social media profiles and linked NFT assets. Our dataset contains 245,159 tweets posted by 17,155 unique users, directly linking 62,997 NFT assets on OpenSea worth 19 Million USD. We have made the dataset public. We analyze the growth of NFTs, characterize the Twitter users promoting NFT assets, and gauge the impact of Twitter features on the virality of an NFT. Further, we investigate the effectiveness of different social media and NFT platform features by experimenting with multiple machine learning and deep learning models to predict an asset's value. Our results show that social media features improve the accuracy by 6% over baseline models that use only NFT platform features. Among social media features, count of user membership lists, number of likes and retweets are important features.
1 INTRODUCTION
The paper investigates whether Twitter activity helps explain NFT value, a difficult task because NFT prices are volatile and sales are sporadic. It links Twitter and OpenSea data and evaluates models using platform and image features.
1 INTRODUCTION
- NFT valuation is challenging because sales are sporadic and prices volatile, making a single sale an unreliable indicator of an asset’s overall value.The paper frames valuation as a social phenomenon involving marketing, recognition, and popularity.Source: §1 · p. 2
- OpenSea is selected as the marketplace for the study because it is described as the largest and most popular NFT marketplace.The passage reports over 300,000 users and $3.4 billion in trading volume in August 2021.Source: §1 · p. 2
- The authors focus on Twitter because it accounts for more than 70% of OpenSea’s social-media traffic and may help build public perception and attract buyers.They cite high-priced NFT sales promoted by well-known personalities as motivation to study social-media influence.Source: §1 · p. 2
- The study collects tweets announcing NFTs, follows their OpenSea links, and analyzes Twitter, OpenSea, and NFT-image features to assess asset value.It motivates average selling price as the valuation metric because sales are limited and prices highly volatile.Source: §1 · p. 2
- The best ensemble of Twitter and OpenSea features achieves 69.5% accuracy in five-class asset-value classification; image features have limited predictive power.The paper reports that both Twitter and OpenSea features influence model output, with branding and metadata having a stronger effect than the NFT image.Source: §1 · p. 3
2 RELATED WORK
Prior NFT research has examined blockchain protocols, market correlations, and fairness, while NFT asset valuation has received limited attention. This paper focuses on valuation using OpenSea and social-media features.
2 RELATED WORK
- Prior NFT work has examined protocols, standards, security, market correlations with crypto-assets, individual collections, and fairness in NFT submarkets.The cited studies include analysis of the broader NFT market and collections such as Axie, CryptoPunks, Decentraland, and CryptoKitties.Source: §2 · p. 3
- NFT asset valuation has limited prior work; Nadini et al. [11] predicted value using sale history and visual features but omitted social-media features.The paper accordingly focuses on the impact of OpenSea and social-media features on valuation.Source: §2 · p. 3
- Social-media features have also been used to predict prices in traditional markets, including stock prices.The paper uses this prior work as context for studying social-media features in NFT valuation.Source: §2 · p. 3
3 THE NFT ASSET VALUATION PROBLEM
Because NFT sales are irregular and sparse, the paper formulates asset valuation rather than periodic price prediction. It defines value from historical average sale price and assigns assets to five logarithmic classes.
3 THE NFT ASSET VALUATION PROBLEM
- Only 0.9% of assets in the dataset had more than five sales, limiting the regular historical observations available for conventional price prediction.NFT markets are described as illiquid, with volatile, irregular sale prices and few transactions per asset.Source: §3 · p. 3
- The paper defines an NFT’s asset value as its average selling price across historic sales, intended to compensate for volatile individual sale prices.The authors use this valuation task instead of price prediction for an illiquid market.Source: §3 · p. 3
- Assets are assigned to five value classes based on average sale price, with class boundaries binned logarithmically because maximum and minimum average prices differ by orders of magnitude.The classes range from Very Low Value Asset to Very High Value Asset and are referred to as Class1 through Class5.Source: §3 · p. 3
4 PRIMER ON OPENSEA PLATFORM
OpenSea is a peer-to-peer NFT marketplace where assets are identified through contract and token information, and can be listed, sold, and resold. Assets may also be grouped into collections, while transactions can incur gas fees.
4 PRIMER ON OPENSEA PLATFORM
- An OpenSea asset is a unique digital item stored as an NFT; its contract address and token ID identify it, while metadata links to its media.Creators list assets for sale at fixed prices, in first-price or Dutch auctions, or through direct offers.Source: §4 · p. 3
- Assets can change owners and prices through resale, and collections group homogeneous assets with shared traits and properties.CryptoPunks is given as an example of a collection whose assets share traits with small variations.Source: §4 · p. 4
- Buying assets and accepting offers on OpenSea can incur an Ethereum transaction gas fee in addition to the asset price.The passage notes that most OpenSea transactions use Ethereum.Source: §4 · p. 4
5 DATA COLLECTION AND ANALYSIS
The study links Twitter activity with OpenSea assets and examines their interaction, user profiles, and asset value. Tweets closely track NFT creation dates, while follower counts show only a weak positive relationship with asset value.
5.1 Data Collection
- The collection pipeline links Twitter and OpenSea data through each asset’s unique OpenSea URL.Source: §5.1 · p. 4
5.1.1 Twitter Data.
- The Twitter collection contains 245,159 tweets from 17,155 users, linked to 62,997 OpenSea assets across 16,001 collections, with tweet and user metadata.Source: §5.1.1 · p. 4
5.1.3 FAIR Dataset Principles.
- OpenSea asset features and images were collected, using the first frame for videos and GIFs; the dataset is public, CSV-formatted, documented, and described as FAIR.The data came from public API endpoints, and the authors report that collection adhered to platform privacy policies.Source: §5.1.3 · p. 4
5.2.1 Correlation between NFT popularity across platforms.
- Spearman’s correlation between daily tweet and NFT-creation counts was 0.85 (p-value < 0.001); 54.6% of tweets followed creation by less than a day.The analysis aggregates features of tweets mentioning each asset.Source: §5.2.1 · p. 4, Figure 4 · p. 4
5.2.2 Delay between posts on Twitter and OpenSea.
- The delay between NFT creation and Twitter mention approximately followed a log-normal distribution, but showed no significant correlation with asset value (ρ < 0.05).More than half of tweets appeared within a day of NFT creation.Source: §5.2.1 · p. 4, §5.2.2 · p. 5, Figure 5 · p. 5
5.2.3 Twitter Username Analysis.
- Over 60% of NFT-affiliated accounts were created in March 2021, versus 18% of non-affiliated accounts; non-affiliated usernames promoted assets with marginally higher mean value (about 25%).The authors attribute the value difference to many low-quality accounts created specifically to push NFT content.Source: §5.2.3 · p. 5, Figure 6 · p. 5
5.2.4 Asset Value Analysis.
- The Spearman coefficient between log-scale follower count and asset value was 0.20, indicating a weak positive correlation.Follower count is the average across users who tweeted about each NFT; values below one were filtered out.Source: §5.2.4 · p. 5, Figure 7 · p. 5
6 FEATURES AND MODELS FOR ASSET VALUATION
The paper trains multiple machine-learning and deep-learning models to predict NFT asset value using Twitter, OpenSea, and image features.
6 FEATURES AND MODELS FOR ASSET VALUATION
- Multiple machine-learning and deep-learning models use Twitter, OpenSea, and image features to predict NFT asset value.Source: §6 · p. 5
6.1 Features
The feature set combines Twitter account- and tweet-level signals with OpenSea asset-level information. Twitter attributes are aggregated across tweets and users, while OpenSea features include asset metadata and marketplace events.
6.1 Features
- The study uses 77 Twitter features and 19 OpenSea features spanning account, tweet, and asset levels.Table 1 summarizes salient features from these groups.Source: §6.1 · p. 5, Table 1 · p. 6
6.1.1 Twitter Features.
- Twitter likes, replies, retweets, and account attributes are aggregated across tweets and users, using mean, maximum, and minimum values.Multiple tweets can mention one asset, and multiple users can mention the same asset.Source: §6.1.1 · pp. 5–6
6.1.2 OpenSea Features.
- OpenSea features include asset creation date and bid or offer events, with counts of sales, bids, and offers derived from marketplace events.Offers and bids indicate purchase intent; accepted offers transfer the asset, while auction bids transfer it to the highest bidder at expiration.Source: §6.1.2 · p. 6
6.2 Problem Settings: Binary vs Ordinal
The paper frames NFT valuation as both a binary profitability task and an ordinal multiclass task. The latter preserves the natural ordering of value intervals rather than treating them as independent labels.
6.2 Problem Settings: Binary vs Ordinal
- NFT asset valuation is modeled as both binary and ordinal multiclass classification.Source: §6.2 · p. 6
6.2.1 Binary Classification.
- 78% of assets were unsold or sold for less than $10 and form the loss-bearing class; assets sold for more than $10 form the profitable class.Source: §6.2.1 · p. 6
6.2.2 Multiclass Ordinal Classification.
- Profitable assets are divided into five logarithmically binned classes, from $10–$100 through higher powers of ten; only 23 of 14,814 assets fall in Class 5.The maximum selling NFT was $552,621.Source: §6.2.2 · p. 6
- Because the five value classes have an intrinsic order, the paper uses ordinal classification with four threshold classifiers predicting P(X > i).It also evaluates ordinal performance with an index where smaller values indicate better classification.Source: §6.2.2 · p. 6
6.3 Models
The paper evaluates traditional machine-learning models using Twitter and OpenSea features and CNNs using NFT images. The image-based models apply the same ordinal classification framework to representative asset images.
6.3 Models
- The study experiments with machine-learning and deep-learning models for NFT valuation.Source: §6.3 · p. 6
6.3.1 Traditional machine learning models.
- Logistic regression, SVM, random forests, LightGBM, and XGBoost are tested with Twitter and OpenSea features for binary and ordinal tasks.Source: §6.3.1 · pp. 6–7
6.3.2 CNNs for Image-based Predictions.
- ResNet-101 and DenseNet-121 CNNs predict value from representative NFT images using four binary classifiers and a final Softmax layer.Training uses image augmentation and batches of 128.Source: §6.3.2 · p. 7
7 ASSET VALUATION RESULTS
Combining Twitter and OpenSea features improves ordinal valuation performance over OpenSea features alone, and XGBoost performs best among the tested classifiers. Image-only models perform worse, while adding image features does not improve the ensemble.
7.2 Accuracy using Different Classifiers
- XGBoost achieves the best ordinal accuracy and lowest ordinal index among the tested classifiers on the Twitter + OpenSea ensemble.The experiments use a 75% training and 25% test split.Source: §7.2 · p. 7, §7 · p. 7, Table 3 · p. 8
7.3 Accuracy with CNNs
- 54.62% and 52.94% ordinal classification accuracy are reported for DenseNet-121 and ResNet-101, respectively, below the Twitter-only and OpenSea-only models.Adding image features to the Twitter–OpenSea ensemble does not improve performance.Source: §7.3 · p. 7
7.1 Accuracy with Different Feature Sets
- Twitter features such as listed count, maximum likes, and replies rank alongside OpenSea offer and bid signals among important ensemble features.The figure identifies both platform feature groups in the importance scores.Source: Figure 8 · p. 7, §7.1 · p. 7
- 69.33% ordinal classification accuracy for Twitter + OpenSea versus 63.22% for OpenSea-only features; the ensemble's ordinal index improves from 0.39 to 0.34.Source: §7.1 · p. 7, Table 2 · p. 8
7.3 Accuracy with CNNs
- The authors interpret the weaker image-only results as consistent with branding and context around an NFT influencing value more than its content.Source: §7.3 · p. 8
8 CONCLUSION
The paper argues that social-media signals, especially Twitter listed counts and username characteristics, affect NFT asset value and could support trading strategies. It also identifies sparse, rapidly evolving markets and manipulated transactions as limits of the current system.
8 CONCLUSION
- Twitter listed counts and username characteristics significantly affect asset value, and the paper presents its work as the first to characterize and value NFTs with social-media features.Source: §8 · p. 8
- The proposed system may support profitable trading strategies by identifying overvalued and undervalued assets.Source: §8 · p. 8
- Because NFT markets remain nascent with limited data, large data-hungry deep-learning models are not viable in this setting.Source: §8 · p. 8
- The current system would not handle outlier transactions and schemes such as self-trading to inflate prices and create social-media virality; future work targets fraud detection.Source: §8 · p. 8