Another month has passed and we’ve finally reached the final form of George! On the same good note, our testnet is live, and feel free to check it out! The most awaited trick in George’s sleeve — pricing any NFT (or Georging any NFT) just based on the contract and asset id is functioning! It has been a steep and challenging path but we got here all in due time. Let’s dive deeper and unpack our approach to ensure we all are on the same page on the inner workings
How do you price any NFT? If we take a peek at what usual features we have for NFTs listed on OpenSea will see attributes (like glasses, hats, and colors). Pricing on them would give a great fit but it would work only within the collection of interest. For example, we price Bored Ape Yacht Club NFTs, the same pricing engine would have difficulties if applied on Doodles. George takes a bit different approach as we use deep AI to uncover the natural features of each NFT as discussed before. Therefore, we are not bound to price only individual collections!
Let’s reframe the problem differently — we have a collection of 90s used cars and only have pictures of them. The prevailing design will be boxy and squarish-looking cars but it’s not listed under the features of the car. George here will observe squarish features and will price them accordingly. Once we take this pricing engine and apply it to a newer-looking car — it will struggle greatly as squarish features are no longer prevailing these days. That is — ignoring the utility and specs of cars themselves. However, if we price all cars in the world in all periods (or in our case — major traded collections of NFTs) then the story is different!
Why is that? Strategically, George leverages Stein’s paradox where instead of building high precision models for each asset type, he is aiming for a broad set of pricing engines with a slight bias in each (going slightly below the value of the asset). But it’ll be closer to the true value 8 out of 10 times as opposed to a highly accurate case with 9 hits and a terrible miss in the last one. This is subject to further improvements as we are pioneering the whole approach here. In technical lingo, we are building on a wide range of weak learners (pricers). Weak learners are algorithms performing slightly better than random guessing on average. Once we have an in-depth grasp of them — we combine and boost them to formulate a strong learner. It’s equivalent to a real-world case of running a survey on an extensive group and aggregating their results. Individually some of them will be horribly off but on average — they will be closer to the real value. As our project grows and coverage collection expands — performance will continue to improve and George will learn from his mistakes for every request you make. To err is human, but to revise is divine!
Let’s get to nitty-gritty details and price Metroverse Block 16# (left subplot below) with the highest offer on OpenSea at 4.3k euros. Note that we don’t have Metroverse collections in the coverage so we are partially giving our 90s used car pricing engine a new Tesla Model S to price. The prices we are getting are widely distributed as we see on the right-side plot. Our prototype is pricing it to be worth around 5k with the median while the latest valuation is at 4.3k — a slight miss as he seems to be pricing when ETH was closer to pre-challenging times.
Let’s unpack it from a different angle — individual pricing engines within each collection. At the moment we have four different methods and it’s up for expansion in the near future. As we see, all of them tend to agree on the 4k-ish valuation and only the first method tends to give a few wide-shot ideas.
For a different perspective, we can unpack them from the individual collection pricers where we clearly see our A pricer providing more extreme deviations. The remaining three are highly similar and compact. This would indicate a high degree of homogeneity and we’ll need to expand with more unique pricers going forward.
On the collection, on dimension more dispersed view emerges. While it’s intuitive from the ML/AI sense as if you have only observed cars from the 90s with similar valuations — it’s hard to predict the price of Tesla now. Still, except for a few costlier collections — most tend to deviate towards the 3kish mark.
Of course, this is an early prototype of things to come but it’s all about incremental progress where we are going from nothing to something. In the following iterations, we’ll get better and more accurate valuations as we’ll expand on collections and pricer angles increasing the width and depth of George. Going forward updates will be irregular and once the evaluator is up online so you’ll be able to George your favorite NFTs!