Inspiration
We want to predict the prices of NFTs, from sparse public bids.
What it does
Predicts Cryptopunk prices, and as a side effect, Punk similarities.
How we built it
Graph attention (neural network) -- based on Punks and two years' bid history, represented as a sparse bipartite graph.
Challenges we ran into
Estimating prices in very skewed and rapidly rising valuations. Some hidden and suspicious bids -- but most are public. Very lucky with that.
Accomplishments that we're proud of
It works! Prices are reasonable (if a bit out of date). We trained a model with no underlying assumptions about what Punk attributes matter, it learns those through the price history!
What we learned
This can work! Also, which Punks are likely to go up in price soon....
What's next for Predicting the Price of CryptoPunks Using ML
Bigger models, more advanced loss functions, live daily prices based on up to the date bids. Computer vision (CV) models on the Punks -- right now we just use named attributes. While some Punks are cool, with that je ne se quoi that can't be represented in an attributes vector.
Inspiration
We want to predict the prices of NFTs, from sparse public bids.
What it does
Predicts Cryptopunk prices, and as a side effect, Punk similarities.
How we built it
Graph attention (neural network) -- based on Punks and two years' bid history, represented as a sparse bipartite graph.
Challenges we ran into
Estimating prices in very skewed and rapidly rising valuations. Some hidden and suspicious bids -- but most are public. Very lucky with that.
Accomplishments that we're proud of
It works! Prices are reasonable (if a bit out of date). We trained a model with no underlying assumptions about what Punk attributes matter, it learns those through the price history!
What we learned
This can work! Also, which Punks are likely to go up in price soon....
What's next for Predicting the Price of CryptoPunks Using ML
Bigger models, more advanced loss functions, live daily prices based on up to the date bids. Computer vision (CV) models on the Punks -- right now we just use named attributes. While some Punks are cool, with that je ne se quoi that can't be represented in an attributes vector.
Log in or sign up for Devpost to join the conversation.