Exploring Gender and Race Biases in the NFT Market

Howard Zhong, Mark HamiltonarXiv 2023 · DOI · arXiv

Introduction

As the NFT market exploded in 2021, concerns arose regarding "digital discrimination." Preliminary reports suggested that female and darker-skinned avatars were selling for significantly less than their male and lighter-skinned counterparts. However, these claims were largely based on visual inspections of price graphs from a single collection: CryptoPunks. This study moves beyond anecdotal evidence by applying formal hypothesis testing to determine if these price disparities are statistically significant and whether they persist across the wider NFT ecosystem.

Methodology

Analyzing NFT prices is statistically challenging because the data is "heavy-tailed"—a few rare NFTs sell for millions, while most sell for very little. To handle this, the authors employed several robust techniques:

  • Log-Transformation: Applied to normalize the power-law distribution of prices, allowing for more accurate Student’s t-tests.
  • Paired t-tests: Used to isolate identity bias from general market volatility by comparing male vs. female prices within the same day or week.
  • Winsorization: Outlier trimming (0.1% to 5%) ensured that extreme "whale" purchases or listing errors did not skew the conclusions.

The study utilized a massive crawl of the OpenSea marketplace, encompassing an initial dataset of 790 collections. From this, the authors derived a gender-labeled dataset of 44 collections and a race-labeled dataset focusing on CryptoPunks, Avastar, and Dynamic Duelers.

Key Insights and Results

Racial Bias is Statistically Significant

The analysis confirmed a significant price premium for lighter-skinned avatars. In CryptoPunks, lighter-skinned avatars consistently commanded higher prices than darker-skinned ones (p < 0.001). This trend was also suggested in other race-labeled collections, though smaller sample sizes made those results less definitive.

Weekly mean sale prices for CryptoPunks by skin color, showing a consistent premium for lighter-skinned avatars.

The authors hypothesize that this bias is driven by a "demographic mirror" effect: buyers often prefer NFTs that look like themselves for use as profile pictures (PFPs). Geographic data shows that NFT interest is heavily concentrated in North America and East Asia, with significantly lower interest in Africa and India. This geographic skew in the buyer pool likely creates higher demand—and thus higher prices—for lighter-skinned avatars.

Challenging the Gender Bias Narrative

Contrary to previous anecdotal reports, the study found no statistically significant price difference between male and female CryptoPunks. Across the broader 44-collection market, the mean price difference was a negligible 0.63%.

Market-wide analysis of gender pricing: (a) weekly mean sale prices and (b) monthly sale price distributions show no significant gap between male and female avatars.

Out of 44 collections, only 11 showed a significant "male premium," while 7 actually showed a "female premium." The remaining 26 showed no significant difference. This lack of gender bias likely reflects a relatively balanced investor pool; data indicates that among the most active age group (18-34), 24% of men and 21% of women own NFTs.

Conclusion and Future Work

The primary limitation of this research is the scarcity of race labels in NFT metadata. While gender could be analyzed across 44 collections, race analysis was limited to just three due to a lack of explicit labeling. Future work aims to use machine learning to automate race labeling for a larger portion of the NFT market. The authors also suggest that "fairness-aware" NFT launches—where identity traits are distributed more equitably or decoupled from rarity—could help mitigate these emerging market biases.

Summary updated

BibTeX
@misc{zhong2023exploring,
  author = {Howard Zhong and Mark Hamilton},
  title = {{Exploring Gender and Race Biases in the NFT Market}},
  year = {2023},
  eprint = {2304.06484},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2304.06484},
}
Figure