NFT Finance Part 2: The pesky challenge of pricing NFTs at scale

Proponents of deep learning say it can be applied to some of society’s most complex problems. Nikolai Yakovenko is now turning it on one of crypto’s: NFT pricing.

Yakovenko, a deep learning expert who has worked for Google, Twitter and Nvidia, has already built deep learning models for poker and sports predictions. Now he’s founded a startup called DeepNFTValue, which is applying those techniques to model the value of NFTs.

While it might sound like a trivial problem — after all, NFTs are bought and sold every day — the lack of a reliable way to determine the value of an NFT is holding back new kinds of financial applications, like NFT-backed loans. 

The reason why such a method doesn’t exist yet is that pricing NFTs is an extraordinarily difficult task. That’s because every collection is unique, individual items from a collection are often highly illiquid, and many NFTs can be extremely valuable but have little-to-no trading history to use as a reference point — like a CryptoPunk that has never been sold. On top of that, NFT purchases are sometimes made purely based on the aesthetic look of the NFT (such as Eminem buying a Bored Ape that looks like him). 

Yakovenko is convinced that deep learning tools will help solve the NFT pricing problem. But not everyone agrees that it is a problem that an algorithm can solve by itself.

Modeling the complex world of NFTs

Yakovenko’s DeepNFTValue currently provides a valuation for all CryptoPunks, which is updated every few days. It shows Punks that are for sale and identifies which ones are on offer below their valuation. It also breaks down attributes, such as hats, hoodies and 3D glasses, by their average price. He says he is planning to add support for Bored Ape Yacht Club NFTs in the near future. 

DeepNFTValue works by tracking all the publicly available variables related to NFTs. This includes the aforementioned attributes of the NFTs, along with previous sales, offers and bids. It also tracks the wallets that most frequently buy and sell CryptoPunks. Sometimes Yakovenko will also add his own synthetic attribute, one not listed in the NFT’s metadata — such as a Punk having a bald head, for example.

It takes all of this data and uses deep learning algorithms to generate what it estimates are relatively accurate prices for each NFT in the collection — with the caveat that the market is changing rapidly and people often decide to make discounted sales.

Yakovenko says the challenge he finds most interesting is “dealing with learning current prevailing prices based on sparse transactions.”

“A little bit of alpha: the thing that matters the most is the offers,” he says. “If you have a piece you’re listing at 120 ETH and nobody buys it — everyone could have taken a shot at it — you know its maximum value is below 120 ETH.”

But offers don’t tell the whole story either. Yakovenko blames “sniping,” a process where a bid is accepted and a bot comes in at the last instant and offers a slightly higher price. While this has been largely solved — traders can now make offers that only a specific person can accept — it can still deter offers.

Yakovenko acknowledges that high transaction fees on Ethereum can deter offers too. “The gas is going to matter a lot more for cheaper collections,” he says.

That’s a lot for a pricing algorithm to have to deal with. “Question is: can you get the model to learn all of that in a systematic non-biased way? In reality, it’s pretty challenging. There’s no playbook for this, this has never been done before,” says Yakovenko.

Ultimately, though, the goal is not to create a tool that investors can use to make bets, he says. “The point is not to make the most accurate predictions. The point is to make reasonable predictions all the time that are hopefully not systematically biased.”

The human factor

No matter how good they are, though, models won’t be successful without human input, says Nick Emmons, co-founder and CEO of NFT valuation platform UpShot. 

Emmons wants to combine algorithmic pricing with human appraisals. UpShot aggregates data from its own internal models, third-party models and a group of human agents — with the idea that the combination should be better at zeroing in on an NFT’s true value.

“We’re trying to build this aggregation mechanism, this substrate, in the effort of producing a single data point — a single price point for as many NFTs as possible across this space,” says Emmons.

At first, Upshot tried using human appraisals as the starting point, but found that this was too inefficient. Humans couldn’t be asked to just name prices for NFTs, they would need to compare two different ones and say which was more valuable — a slow process that only produces rough price valuations. 

Instead, Upshot currently uses machine learning models to value 220,000 NFTs each hour. But it still thinks human appraisals are valuable. “Human appraisal is much more effective as a supplemental device,” says Emmons.

Upshot’s rationale is that NFTs can be priced even without market interactions. So if something hasn’t been traded for a while, a human can still help provide some indication of the rough value of the NFT based on their view on the market and other collections. The model also uses sentiment analysis from Discord and Twitter.

Human appraisals do come with a challenge, though. For example, participants may have reason to lie — say, if they own lots of NFTs in a particular collection and want to inflate its value. Emmons says that Upshot incentivizes people to respond honestly and analyses the responses to check they are accurate. 

Emmons says a lack of price discovery mechanisms and liquidity are limiting the applications of NFTs. But if traders are able to open up positions with much more ease, that will open up a much wider range of use-cases.

“We’ll see much more NFTs be used for financial assets: real estate, commodities, bonds. The rest that can’t be presented as fungible tokens will be represented as NFTs,” he says. “We’ll be able to realize the full picture of what’s possible with DeFi.”

This is the second article in a three-part series focused on the financial applications of NFTs. Read Part 1 here.