Non-fungible tokens, or NFTs, have been around for several years, but interest in them has rapidly increased in 2021 with a number of high-profile sales, including the $69.3 million Christie’s sale of the NFT associated with Everydays: the First 5000 Days – a digital artwork by the US artist Beeple. Despite concerns about the environmental impact of the underlying technology, NFTs racked up sales of $2.5 billion in the first half of 2021 – an all-time high. But why do some NFTs sell for so much? That question is at the heart of a new Turing-funded study (part of the Turing’s token economy theme) that we recently published in Scientific Reports – the first comprehensive, quantitative overview of the NFT market.
First, the basics. An NFT is a unique unit of data that can be used to certify the ownership of a digital asset. NFTs are created using blockchain technology – usually the Ethereum blockchain – and they can represent any digital object, including photos, artworks, memes, music, digital collectibles and even tweets. There are many NFT marketplaces, where sellers either sell through an auction or a ‘buy now’ option. The novelty of NFTs compared to ‘physical’ art objects is that digital assets can in theory be replicated and/or downloaded an unlimited number of times, and can therefore be ‘owned’ by anyone. The idea behind NFTs is therefore to establish which of the many copies is the ‘original’, and who owns it.
To map the evolution of the NFT market, we looked at a dataset of 4.7 million NFTs exchanged by more than 500,000 buyers and sellers between 23 June 2017 and 27 April 2021. Items included digital art as well as other collectibles such as digital cards and items used in video games (e.g. weapons and clothing). There was a large variability in the prices of the analysed NFTs: only the top 1% of objects traded for more than $1,500, with 75% of objects selling for less than $15.
In order to find out what determines the sale price of an NFT, we developed a machine learning model that considers three factors: 1) visual characteristics, 2) previous sales of related NFTs, and 3) the popularity of the buyers and sellers.
Visual features of the digital assets were assessed using sophisticated computer vision techniques that extract the visual properties of the objects, such as their colours and shapes, and look for similarities. Previous sales were assessed by considering the market history of items from the same collection (items exchanged on the NFT market are organised in collections that, in most cases, share some common features). Finally, the popularity of traders in the NFT network was assessed by considering the total number of purchases and sales made by each trader.
We found that all three factors play a role in determining the price of an NFT. Previous sales of related NFTs are consistently the most important feature, explaining up to 50% of the variability in NFT prices. Visual features are also important determinants of the price, increasing the performance of our machine learning model by 10-20%. Adding data about the popularity of traders further increases the performance of our model by 10%. Together, these three factors can explain up to 70% of the variability in NFT prices.
These results suggest that it might be possible to accurately predict NFT prices – a direction we are planning to explore. We have also made our data freely available so that others can perform their own analysis. Better price prediction would help to encourage a more mature financial industry around NFTs, reducing investment risks connected to the unpredictable volatility of NFTs, both for private investors and at the systemic level.
NFTs are revolutionising the way in which digital content is produced and exchanged. Our study will stimulate further research on NFT production, adoption and trading across a broad array of disciplines, including economics, law, cultural evolution, art history, computational social science and computer science.
Video and top image created by Mauro Martino