PaperlayerView exact revision on arXiv

Source-linked AI summary

Mapping the NFT revolution: market trends, trade networks and visual features

Matthieu Nadini, Laura Alessandretti, Flavio Di Giacinto, Mauro Martino, Luca Maria Aiello, Andrea Baronchelli

arXiv:2106.00647v4q-fin.STcs.CYphysics.soc-ph

TL;DR

Research on NFTs has largely focused on technical aspects, leaving the market’s broader structure and evolution insufficiently characterized. Using millions of trades, the paper maps market properties, trader and NFT networks, visual homogeneity, and price predictability, finding specialized traders, homogeneous collections, and strong predictive value from sale history and visual features.

  • Problem

    NFT research remains limited and focuses mostly on technical aspects, leaving broader market structure and evolution insufficiently characterized.

  • Method

    The paper analyzes 6.1 million trades of 4.7 million NFTs, characterizing market properties, interaction networks, visual features, and price predictability.

  • Results

    Traders are mostly specialized, NFT collections tend to be visually homogeneous, and sale history plus visual features help predict NFT prices.

  • Takeaways & Limitations

    The study provides an initial overview of NFT markets across categories and identifies trading specialization, collection homogeneity, and price-predictive NFT properties.

  • Takeaways & Limitations

    The authors identify limitations of the study as directions for further research.

Abstract

from arXiv · show

Non Fungible Tokens (NFTs) are digital assets that represent objects like art, collectible, and in-game items. They are traded online, often with cryptocurrency, and are generally encoded within smart contracts on a blockchain. Public attention towards NFTs has exploded in 2021, when their market has experienced record sales, but little is known about the overall structure and evolution of its market. Here, we analyse data concerning 6.1 million trades of 4.7 million NFTs between June 23, 2017 and April 27, 2021, obtained primarily from Ethereum and WAX blockchains. First, we characterize statistical properties of the market. Second, we build the network of interactions, show that traders typically specialize on NFTs associated with similar objects and form tight clusters with other traders that exchange the same kind of objects. Third, we cluster objects associated to NFTs according to their visual features and show that collections contain visually homogeneous objects. Finally, we investigate the predictability of NFT sales using simple machine learning algorithms and find that sale history and, secondarily, visual features are good predictors for price. We anticipate that these findings will stimulate further research on NFT production, adoption, and trading in different contexts.

I. INTRODUCTION

NFT adoption and trading expanded rapidly, while comprehensive empirical understanding of the market remained limited. This paper addresses that gap by analyzing millions of trades across blockchains, trader and asset networks, visual features, and sale-price predictability.

  • Motivation: $2 billion: NFT trading volume in the first four months of 2021 exceeded the entire 2020 volume tenfold.NFT adoption also expanded beyond art into gaming, music, video, and other digital-content industries.
  • Background: NFTs are blockchain-stored data units certifying unique, non-interchangeable digital assets and their ownership provenance.They can represent photos, videos, audio, art, gaming objects, and sports collectibles across multiple blockchains.
  • Research gap: Research on NFTs remained limited, emphasizing technical issues and studies of individual collections or marketplaces rather than the overall market.Prior findings included declining values from digital abundance, cryptocurrency-linked prices, speculation, expert effects, and centralized co-ownership networks.
  • Study design: 6.1 million trades of 4.7 million NFTs: the study provides a comprehensive quantitative overview covering 160 cryptocurrencies, primarily Ethereum and WAX.The dataset spans June 23, 2017 to April 27, 2021 and supports analyses of market evolution, trader and asset networks, visual features, and predictive models.
  • Scope and limitation: NFT categories are manually classified, but assigning objects to categories is outside the paper’s scope because Art, Collectibles, and Game objects can overlap.The classification references NonFungible Corporation and OpenSea while acknowledging ambiguous boundaries between categories.

II. RESULTS · A. The NFT market

The NFT market remained near $60,000 in daily volume until mid-2020, then expanded dramatically to over $10 million daily by March 2021. Market composition, prices, trading frequency, collection sizes, and secondary-sale dynamics varied substantially across categories and collections.

  • A. The NFT market: After CryptoKitties’ late-2017 growth, daily trading averaged ∼60 000 US dollars until mid-2020 before surpassing ∼10 million US dollars in March 2021.The March 2021 volume was 150 times larger than eight months earlier.
  • A. The NFT market: Until the end of 2018, Art dominated the market, while from January 2019 other categories gained popularity in volume and transaction counts.Between January 2019 and July 2020, Art, Games, and Metaverse together represented ∼90% of exchanged volume, contributing 18%, 33%, and 39%.
  • A. The NFT market: Since mid-July 2020, Art contributed ∼71% of transaction volume and Collectible assets 12%, whereas Games and Collectible represented 44% and 38% of transactions.Only 10% of transactions involved Art, indicating higher average Art prices than other categories.
  • A. The NFT market: NFT prices were broadly distributed: 75% of assets averaged below 15 dollars, while 1% averaged above 1 594 dollars.The top 1% averaged above 6 290 dollars for Art, 9 485 dollars for Metaverse, and 12 756 dollars for Utility; four Art NFTs exceeded 1 million dollars.
  • A. The NFT market: Only ∼20% of assets had secondary sales, and the tail for assets with s ≥10 sales followed P(s) ∼s^β with β = −1.4.Only 0.07% of all assets were sold more than 10 times; some Games and Art assets were sold more than a thousand and five thousands times.
  • A. The NFT market: Collection sizes followed P(n) ∼n^α with α = −1.5, with ∼75% of collections containing fewer than 37 unique assets and ∼1% exceeding 10 400.The broad size distribution contrasts small collections with a small number of very large collections.
  • A. The NFT market: Secondary-sale price patterns were collection-specific: CryptoKitties initially sold below primary prices, while prices rose in 2021 as potential customers increased.Alien alternated between rising and falling secondary prices, whereas Unstoppable secondary sales were rare because its NFTs were blockchain-secured web domains.
  • A. The NFT market: Secondary prices were lower than primary prices in 66% of cases in 2017, but only 27% of cases in 2021.This shift accompanied the changing secondary-sale patterns observed across collections.

B. The networks of NFT trades

The NFT trade network is highly heterogeneous and collection-specialized: a small fraction of traders accounts for most activity, while traders tend to transact with others focused on the same collections. The NFT network is strongly clustered by collection but still contains cross-collection purchase sequences and large connected components.

  • Trader network: 85% of all transactions are performed by the top 10% of traders, who trade at least 97% of all assets.Trader strength follows a power-law distribution with exponent λ1 = −1.85.
  • Trader network: At least 73% of traders’ transactions occur in their top collection and at least 82% in their top two collections combined.The strongest specialization is found among traders with fewer than ten or more than ten thousands of transactions.
  • Trader network: Q = 0.613 for the collection partition, versus Q = 0.0823 ± 0.0001 for a random network, showing that collection-specialized traders tend to trade with one another.The assortativity coefficient is r = −0.024, close to zero, so traders do not connect based on similar connection patterns.
  • NFT network: NFTs in small collections tend to be bought in sequence with NFTs from other collections, whereas NFTs in large collections tend to be bought within the same collection.NFT strength follows a power-law distribution with exponent λ3 = −3.21.

C. Visual features

Visual features reveal strong graphical homogeneity within NFT collections, with some related collections forming shared visual clusters. AlexNet features reduced by PCA also separate objects by category and provide inputs for sales prediction.

  • Feature extraction and prediction: AlexNet extracts 4 096-dimensional visual representations, whose PC1 to PC5 scores are used to test visual features for predicting sales.Cosine distance ranges from zero for identical images to one for highly different images.
  • Collection-level visual similarity: Within-collection cosine distance is lower than between-collection distance, confirming strong intra-collection graphical homogeneity.The averages are µ = 0.59, σ = 0.20 within collections and µ = 0.87, σ = 0.06 across collections.
  • Collection-level visual similarity: Most collections have distinctive visual styles, with low intra-collection cosine distances for Sorare (CD = 0.24) and Cryptopunks (CD = 0.33), but not Rarible (CD = 0.89).Pixel-art collections such as Chubbie, Cryptopunks, and Wrapped Punks share visual features, as do Cryptokitties and Axie.
  • Dimensionality reduction: The first five principal components explain 38.3% of total variance in about 1.25 million graphical objects.Variance is distributed across PC1 to PC5 as 20.3%, 7.3%, 4.0%, 3.8% and 2.7%.
  • Category-level visual similarity: Objects from different categories are separated in the PC1, PC2, PC3 space, with average distance 1.67 bigger than for objects within the same category.Inter-collection distance is 3.17 times bigger than intra-collection distance, and category separation is mainly driven by collection homogeneity.

D. Predicting sales

NFT sale prices are most predictable from prior sales within the same collection, while trader-network centrality and visual features add predictive power. Predicting whether an NFT resells is most accurate for Art NFTs and less reliable for other categories.

  • D. Predicting sales: The median sale price within an NFT collection predicts more than half the variance of future primary and secondary sale prices.Recent sale history is more informative than the entire pre-sale history; a one-week window performs better.
  • D. Predicting sales: R^2_adj declines from 0.90 for predicting next-week median secondary prices to 0.77 for predictions two years ahead.Secondary sale prices are also strongly correlated with prior primary sale prices, but predictive power weakens over longer horizons.
  • D. Predicting sales: Resale occurrence is predicted most accurately for Art NFTs, with F1 > 0.8, versus F1 ∈[0.14, 0.33] for other categories.Resales are uncommon: less than 10% occur within one week of primary sale and about 22% within one year.

III. CONCLUSION

The study provides an overview of NFT markets, finding trader specialization, visually homogeneous collections, and price predictability driven primarily by past history and also visual features. Its limitations identify data, categorization, methodological, platform, attention, and creator-information gaps that motivate future research and marketplace design.

  • Contributions: The study analyzes 6.1 million NFT trades across six categories and examines market properties, trader–NFT networks, collection visual features, and price predictability.The categories include art, games, and collectibles.
  • Results: Most traders are specialized, NFT collections tend to be visually homogeneous, and visual features improve price predictability beyond past sales history.Past history is described as the best predictor, while NFT-specific properties such as associated-object visual features also help.
  • Limitations: The dataset likely misses independent NFT producers, relies on partly arbitrary categorization, and focuses mostly on Ethereum and WAX rather than all NFT platforms.The study gathered data from marketplaces instead of directly from blockchains and notes that other platforms offer smart contracts and NFTs.
  • Limitations: The study did not extensively explore alternative feature extraction, clustering, price-prediction, or market-modeling methods, nor collective attention from social media or Wikipedia.These omissions are presented as directions for future work.
  • Limitations: Price prediction excluded creator information, although creator identity may be important in some contexts, particularly art.Creator identity may be unavailable or nonexistent, including for AI-generated images.
  • Implications: The authors expect the study to stimulate research across disciplines and inform more efficient marketplaces and associated regulation.The proposed disciplines include economics, law, cultural evolution, art history, computational social science, and computer science.

A. Data and methods

The section briefly introduces the study’s data collection and directs readers to Section S2 for detailed methodological descriptions.

  • The study’s data collection is summarized in the following material.
  • The section therefore separates a summary of data collection from the detailed methodological account.
  • Detailed descriptions of the data manipulations are provided in Section S2.

1. Sales data collection · 2. Image collection and visual feature extraction

The study constructs a cleaned NFT sales dataset from blockchain and API records, then extracts visual representations for most associated digital objects using AlexNet and PCA. The resulting data comprise 6.1 million transactions involving 4.7 million NFTs and visual features for about 1.2 million graphical objects.

  • 1. Sales data collection: The dataset includes only NFT purchases involving ownership changes, excluding minting transactions and auction bids.The collection tracks transactions across different cryptocurrencies and blockchains.
  • 1. Sales data collection: Sales records combine data shared by NonFungible Corporation, four Ethereum APIs, and Atomic API monitoring of the WAX blockchain.Ethereum sources include SuperRare, Makersplace, Knownorigin, Cryptopunks, Asyncart, CryptoKitties, Gods-Unchained, Decentraland, and OpenSea.
  • 1. Sales data collection: 6.1 million transactions involving 4.7 million NFTs grouped in 4,624 collections form the cleaned sales dataset.The dataset contains 935 million USD traded.
  • 1. Sales data collection: 160 different cryptocurrencies appear in the dataset, with WAX accounting for 52% of transactions and ETH accounting for 81% of total volume.Most transactions were made in WAX, while USD volume was mostly ETH.
  • 2. Image collection and visual feature extraction: About 1.2 million unique graphical objects were collected for 4.7 million unique NFTs, with at least one object URL obtained for all but less than 3 000 NFTs.The collection focused on image formats and GIFs, and a single digital object could correspond to multiple NFTs.
  • 2. Image collection and visual feature extraction: Animated GIFs were converted to PNGs by extracting their central frame so the visual-feature algorithm could process static images.This conversion preceded neural-network encoding of the images.
  • 2. Image collection and visual feature extraction: AlexNet pretrained on ImageNet encoded each image into a 4 096-value dense vector, which PCA reduced to the 5 most relevant components.The vectors support similarity ranking, clustering, or classification, while PCA preserves data variation as much as possible.

Supplementary Information · S1. ADDITIONAL ANALYSES · S2. ADDITIONAL INFORMATION ON THE DATA AND METHODS

The supplementary analyses extend the study with time-resolved market statistics, network diagnostics, category and blockchain breakdowns, price-prediction analyses, and secondary-sale classification tasks. They also report an overall dataset total of 359,561 buyers, 314,439 sellers, 4,704,479 NFTs, and USD 935.11 million in volume.

  • S1. ADDITIONAL ANALYSES: FIG. S1 tracks unique traders, collections, and transactions over time across categories using rolling 30-day windows.The time series are computed over a rolling window of 30 days.
  • S1. ADDITIONAL ANALYSES: FIG. S2 characterizes buyer–seller transaction weights, trader strength, activity duration, and concentration in top NFT collections.The diagnostics include probability distributions and relationships between strength and the number of active days.
  • S1. ADDITIONAL ANALYSES: FIGS. S3 and S4 repeat network diagnostics by NFT category and blockchain, including cross-collection transactions and largest strong connected components.Panels report average values with bands representing the 95% confidence interval.
  • S1. ADDITIONAL ANALYSES: The supplementary regressions predict primary and secondary sale prices from collection histories, primary-sale prices, and prior collection medians over one-week-to-two-year windows.Results are broken down by NFT categories.
  • S1. ADDITIONAL ANALYSES: Additional analyses examine whether NFTs receive secondary sales within one year and across varying post-primary-sale time windows using F1-score classification.The models use different feature sets or all available features, with results broken down by NFT categories.
  • S2. ADDITIONAL INFORMATION ON THE DATA AND METHODS: 359,561 buyers, 314,439 sellers, and 4,704,479 NFTs generated total volume of 935.11 million USD in the supplementary category breakdown.These totals are reported in the overall category row.

A. Data cleaning and categorization · B. Generation of the traders and NFTs networks of interaction

The study cleans NFT transaction data, categorizes collections, and constructs trader and NFT interaction networks from sequential purchases by buyers. The resulting NFT network includes 4,657,713 of 4,704,479 NFTs, excluding those whose buyers made only one transaction.

  • A. Data cleaning and categorization: Collection names are standardized by removing digits, special characters, unusual patterns, and capitalization, then consolidating variants and assigning generic names to Miscellanea.For example, collections beginning with the Aavegotchi string are renamed Aavegotchi.
  • A. Data cleaning and categorization: NFT transactions are retained using buyer and seller addresses, transaction time, collection name, NFT ID, digital-object URL, cryptocurrency type, and amount.Transactions missing any listed field, except the digital-object URL, are removed.
  • A. Data cleaning and categorization: The analysis computes transaction prices in USD from the cryptocurrency exchange rate on the transaction day and uses addresses as proxies for identities.An individual may have multiple addresses or usernames, so these proxies do not necessarily represent unique real identities.
  • A. Data cleaning and categorization: NFTs sharing common features are grouped into collections, which are assigned to Art, Collectible, Games, Metaverse, Utility, or Other.A collection may belong to more than one category, but forcing each collection into one category is identified as a limitation.
  • A. Data cleaning and categorization: When datasets overlap, duplicated transactions are removed after merging sources, with source-specific priority used to determine which records are retained.The downloaded datasets are not independent, except for the Atomic API.
  • B. Generation of the traders and NFTs networks of interaction: The NFT network links NFTs purchased sequentially by the same buyer, drawing a directed edge from each earlier NFT to the later NFT at the later purchase time.If purchases occur at the same instant, links are drawn from the earlier NFT to each simultaneously purchased NFT.
  • B. Generation of the traders and NFTs networks of interaction: 4,657,713 of 4,704,479 NFTs are included in the constructed network; excluded NFTs belong to buyers who performed only one transaction.The trader network is obtained directly from the collected data, while the NFT network is generated from it using three link-creation rules.

C. Generation of random networks · D. NFT features · E. Sale price regression

The paper constructs degree-preserving random trader and NFT networks, represents NFTs with 11 network, visual, and sale-history features, and uses ordinary least squares regression to estimate sale prices. Features are transformed and scaled before regression, which uses adjusted R2 for evaluation and excludes NFTs without qualifying secondary-sale observations.

  • C. Generation of random networks: Random networks preserve each node’s outgoing and incoming strength by repeatedly swapping randomly selected links from a weight-repeated link pool.The procedure creates 100 independent realizations for both the trader and NFT networks.
  • C. Generation of random networks: The randomization procedure generates 100 independent trader-network realizations and 100 independent NFT-network realizations.Swapping is repeated a number of times equal to the network’s total links.
  • D. NFT features: NFTs are described by 11 features spanning four seller-and-buyer centralities, five AlexNet PCA components, and two collection sale-history measures.The centralities are degree centrality (k) and PageRank centrality (PR); sale-history measures include median prior prices and presale secondary-sale probability.
  • D. NFT features: Using 2 to 10 visual components changed results only slightly: fewer components weakly reduced regression and prediction quality, while additional components added no predictive power.The selected visual features are PCA1...5 components extracted from the object’s AlexNet vector.
  • D. NFT features: For the first NFT sold in a collection, presale secondary-sale probability is set to 0.5; otherwise it is calculated from prior collection sales.For large collections, the probability approaches p (n →+∞) = s/n.
  • E. Sale price regression: Ordinary Least Squares regression estimates primary or median secondary sale prices from NFT features by minimizing the sum of squared residuals.Secondary-price analyses use a time window beginning at ts, with window lengths ranging from one week to two years.
  • E. Sale price regression: Regression excludes NFTs sold too recently for equal-length secondary-sale windows and includes only NFTs with at least one secondary sale in the considered window.Fit quality is evaluated with adjusted R2, which discounts increases caused by adding independent variables.

F. Secondary sale prediction

The study predicts whether NFTs receive a secondary sale using supervised classification trained on earlier primary-sale data and tested on later NFTs. It uses AdaBoost with class balancing and evaluates predictions using F1-score and AUC.

  • Task definition: The target label is 1 when an NFT is transacted in a secondary sale and 0 otherwise.The task predicts resale after the NFT’s primary sale at time ts.
  • Temporal evaluation: 95% of NFTs are used for training and the latest 5% for testing after sorting by primary-sale time ts.This split emulates prediction on future data from past knowledge.
  • Model: AdaBoost is initialized with 100 decision tree stumps and a learning rate of 1.The classifier combines weak learners through a weighted sum.
  • Class balancing: 80% of NFTs are not resold, so random oversampling duplicates positive training samples until the classes are balanced.This increases the model’s effective importance for positive samples without generating synthetic data points.
  • Evaluation: Performance is measured with F1-score and Area Under the ROC Curve (AUC).F1-score combines precision and recall, while AUC assesses confidence-based ranking; AUC is 0.5 for random classification and 1 for perfect ranking.
Loading 2106.00647v4…