TweetBoost: Influence of Social Media on NFT Valuation

Arnav Kapoor¹, Dipanwita Guhathakurta¹, Mehul Mathur¹, Rupanshu Yadav², Manish Gupta¹, Ponnurangam Kumaraguru¹

IIITH¹, IIITD²

Introduction & Motivation

• Blockchain technology has led to products and services that have transformed the financial ecosystem
• NFTs have grown exponentially in 2021 with market cap reaching upto $10 billion
• ‘Everydays: The First 5000 Days’, an NFT artwork by artist Beeple, sold for $69 million.
• Jack Dorsey’s first tweet raised $2.9 million
• NFTs are entirely digital and can be downloaded and shared publicly for free.
• The value of an NFT is based on the perception of buyers
• We use the largest NFT platform opensea and use twitter to understand the influence of social media on the NFT market

Research Questions

• What is the relationship between user activity on Twitter and price on OpenSea?
• Can we predict NFT value using signals obtained from Twitter and OpenSea; and identify which features have the greatest impact on prediction?

Contributions

• We create the first ever data set for NFTs from OpenSea and their corresponding tweets. We have made the dataset public in adherence to the FAIR (Findable, Accessible, Interoperable, Reusable) principles
• We build ordinal classification models to predict NFT asset value using features from both OpenSea and Twitter
• We show that both Twitter and OpenSea features influence the model output. In contrast the predictive power of image features is limited

​

Related Work

• Most of the previous work on NFTs is on analysis of relationship between blockchain, crypto-currency and NFTs
• Previous work on NFT valuation uses machine learning algorithms to predict NFT asset value, social media features are not used
• Social media features have helped predict the prices in stock markets

NFT Asset Valuation

• NFT markets are highly illiquid in nature
• A traditional price prediction setting is not feasible
• In our dataset as well, a very small percentage (0.9%) of the total assets had more than 5 sales
• We define asset value as the average selling price of an asset over all its historic sales
• The asset value is used to divide NFTs into 5 classes

Opensea Primer

• Every NFT is listed as an ‘asset’ and is part of a ‘collection’
• Asset is uniquely identified by an address and token-id
• The owner can list & auctions NFTs
• The buyer can participate in listings and auctions, or directly make an offer
• Most transactions are done on the ethereum blockchain and require a transaction fee

Data Collection

• Twitter Data
• We curate 245,159 tweets from Jan 1, 2021, to Mar 30, 2021, containing opensea.io NFT asset link
• The data set contains 17,155 unique users, 62,997 unique OpenSea assets belonging to 16,001 unique collections
• Opensea Data
• We used opensea API to collect asset and collection level data for NFTs present in Twitter Dataset

Interaction Analysis (P1)

Interaction Analysis (p 2)

Models

• For Twitter-Opensea data
• Logistic regression
• SVM
• Random Forests
• LightGBM
• XGBoost
• For Image data
• ResNet-101
• DenseNet-121
• We divided the data in different feature sets and studied the accuracy separately

Model Results - Feature Sets

Model Results - Twitter + Opensea

Model Results - CNNs

• We obtained accuracies of 54.62% and 52.94% with pretrained DenseNet-121 and ResNet-101 models respectively
• Image models show lower accuracies and F1-scores and higher ordinal classification index values
• On adding image features to the Twitter-OpenSea ensemble, we notice no improvement in model performance

Conclusion

• Track the growth of NFTs and show how social media reach can impact its value.
• Working model to guide sellers about the valuation of their assets
• First work to characterise and value NFT assets using social media features.