Predicting CryptoPunk Prices: The Case for JPEGs

Small pixels, big heart, can’t lose.
Jay-Z, Gary Vee, Steve Aoki and many other celebrities own and display Cryptopunks

Pricing unique & rarely traded assets

Valuing NFTs is difficult, although not also not rocket science. The most valuable and popular NFT project — Cryptopunks — we’ll get into why it deserves its OG status later— minted 10,000 “CryptoPunk” JPEGs back in 2017. Starting a trend that’s been imitated many times since, the JPEGs are “computer generated” with each one unique, distinct a number of distinct enumerated attributes.

Teaching common sense, to a robot

After talking up Punks ever since those Tech Week Uber rides, I put by REPL where my mouth is, and coded up a machine learning model to estimate the daily values of all Cryptopunks.

  • We use an inequality loss, considering both recent bids (minimum) and recent/live offers (maximum). A sale is where bids and offers meet.
  • On the DeepNFTValue website, we show nearest neighbor similarity scores for all Punks. These are derived from the cosine distance of the learned Punk embeddings, based on their Attributes. We add a training loss to encourage Punk price proximity to their neighbors.
  • We train a completely bid-less model to predict prices, only on the Attributes. This model is disconnected from the bid-aware model, but both are encouraged to imitate the output of the other.
  • We don’t actually have a single graph, from which to predict tomorrow’s prices. We have daily snapshots, going back months, and can train the model to predict the next-day sales on all of those. I know this bleeds information between days. Trust me, it helps.
  • Learn long term relationships between Punk attributes and prices, including on a relative basis. If Top Hats and Cowboy Hats are both deemed positive Attributes, mutually exclusive, and worth about the same, such a relationship would persist even on days we had no bids on Cowboy Hats, and as the market is going up (or down?) overall.
  • Reasonable estimates or Punks that never sell. The estimate needs to be grounded in comps, and overall market levels, not free to float entirely — even when there’s no sales, bids or offers from which to triangulate a price.
  • Track the value of Punks and Attributes, over time. Graphs coming soon!

Price observations — August 31, 2021

The run-up in Punk prices ever since Visa bought and displayed a “floor Punk” on Monday — really has “swept the floor” so to speak.

Cryptopunk values — tight band

Hoodies vs Cowboys

Sadly for my friend, a Cowboy Hat enthusiast (contrarian move for a poker player), Hoodies are emerging as the most valuable Human attribute.

Don’t cancel the Punks

Going into this exercise, I was concerned we may discover the market to be biased against darker-faced Human Punks. While we’ve not studied this thoroughly, I think you can browse the price estimates, or recent sales, and notice that this is broadly not the case.

Future Projects

Our immediate goals are:

  • keep updating the website with current prices
  • start running the model back in time and showing some historical values


This is not investment advice. Our estimates are for entertainment purposes only. Please ape responsibly. Only bid on projects that you like, with ETH that you can afford to part with.



Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store
Nikolai Yakovenko

Nikolai Yakovenko


AI (deep learning) researcher. Moscow → NYC → Bay Area -> Miami