MetaQuants: A Solution to Price Discovery for NFTs | by MetaQuants

MetaQuants is the best in class technology for institutional-grade valuations and risk management of NFTs.

Driven by our conviction in the narrative of NFT financialization, we are building the analytics infrastructure for the future of finance.

Our flagship products include an NFT Pricing Algorithm, a Collection Floor Price Oracle, as well as an Analytics and Risk Management Platform.

MetaQuants was born after a hackathon organized by the crypto lending platform Nexo. The goal of the event was to bring talented machine learning and math specialist together, so they can test their skills by figuring a way to appraise NFTs. For a period of 48 hours, participants had to utilize crypto, NFT and sentiment data in order to get closer to demistifying the task.

Our co-founders managed to come up with the most accurate machine learning model, as well as the best overall solution of the problem. The combination of domain knowledge and computing proficiency was key for their success. Their performance led to Nexo providing them with a grant in order to continue with the research and develpment of the product.

Eventually, they managed to built a Pricing Algorithm that could serve as a gateway to the emergence of other verticals in the domain of NFT financialization. In a period of 6 months, our team managed to get from an idea to a product that has the potential to become the data and analytics infrastructure for the future of finance.

There are several inefficiencies regarding the evaluation of NFTs — lack of transparency, overpricing of assets, and inability to choose risk parameters.

In the following paragraphs, the methodology for tackling each problem is defined, and empirical results corroborating the utilized approach are provided.

When it comes to a machine learning pipeline, the feature selection process is as important as the choice of a model. While most existing solutions only include traits of an asset, MetaQuants takes into account:

Macroeconomic Conditions — Proxy Variables:

DJIA
S&P500
Feature Engineering on S&P500 and DJIA in order to extract insights from the raw data

2. Industry Conditions — Proxy Variables:

Bitcoin
Ethereum
Feature Engineering on Bitcoin and Ethereum in order to extract insights from the raw data

3. NFT Collection Conditions:

Feature Engineering which is based on indicators of traditional financial markets, for example, price volatility
Feature Engineering which is based on domain-specific indicators, for example, the floor price of the collection

4. NFT-Specific Conditions:

Rarity Score
Computation of new features which account for past performance of individual assets, for example, the average price of previous sales

After all metrics are generated, Forward Variable Selection method quantifies the predictive power of each indicator, excluding all redundant variables from the optimal model.

Regardless of the viewpoint — lender, borrower, or buyer, overpricing is seen as a focal issue for price discovery algorithms. The idea of purchasing an asset above its fair value hinders the wide adoption of such solutions. MetaQuants employs a model with a specifically defined loss function for the purpose of penalizing overpredictions more than underpredictions.

The following experiment supports this theoretical statement and compares MetaQuants’ model to NFTBank, as a benchmark of the industry’s ability to price non-fungible assets. The selection of a comparison is based on a company’s market share, number of third-party implementations, and seed round valuation.

Note, this is no attempt to discredit NFTBank’s service in any conceivable way. Models are weighted on a statistical basis and there are not any qualitative remarks.

The experiment is structured into a train-test split where the training data is collected from 12th of January to 31st of July and test data covers the period from 1st of August to 3rd of October. Factors for choosing the testing span include number of samples and distribution of the set. There are 889 sales taken from OpenSea, LooksRare, X2Y2, Rarible, and Sudoswap. Other marketplaces and decentralized apps are not considered due to difficulties with data indexing.

NFTBank’s inference is ingested via their API. The historical valuation endpoint returns prices on a daily basis, so we use the value of an NFT at the latest available timestamp before the sale. The results can be observed in Figure 1. MetaQuants’ model, MAPE=15.2%, outperforms its counterparty, MAPE=16.8%, by 150 basis points. As evident from the figure, the two models yield similar results, with MetaQuants’ model having 20% fewer overpriced instances than NFTBank’s model.

Figure 1: Ground Truth Labels versus Predicted Labels

Evaluation metrics are indicators of a model’s ability to predict unseen data. A single figure, however, brings clarity neither to the algorithm’s feature selection nor to its computational approach. As already stated, existing evaluation services function as a black box which undermines them, regardless of their performance.

MetaQuants enumerates the factors for determining a token’s value and specifies the impact of each variable in terms of ETH, Figure 2. The first number, 15.48, represents the average price of all instances in the train set and the value in the top-right corner represents the predicted price for an NFT. Different assets have different feature importance, established by their explanatory variables.

Figure 2: Model’s Explainability (All Values are in ETH)

There is a wide variety of buyer profiles within the NFT ecosystem — from firm believers in the technology to the so-called ”degens”. Apart from a point estimate, MetaQuants provides a value range based on the model’s residuals, so each user can accommodate their individual risk appetite. Risk-averse market participants and lending protocols can choose the lower bound of the interval to protect themselves against overpricing.

Oppositely, bullish traders may select the upper bound for a possible opportunity. The range is dynamically adjusted and accounts for rarity, so different assets have different interval widths. Further, α, can be chosen for a (1 − α) confidence interval, Figure 3. Lower α corresponds to a wider price scope, but also to an additional level of insurance.

Figure 3: Confidence Intervals at α = 0.05 (50 Randomly Selected Samples)

Tackling the problem of price discovery is a stepping stone for the NFT ecosystem on the path to mass adoption. MetaQuants provides a one-stop solution which includes price derivation, feature importance, and confidence intervals for different risk tolerance levels. The company constantly collects feedback from the community on evaluation-related topics, so any emerging concerns can be addressed in a timely manner. Resources are dedicated to research on sentiment analysis, so the model can include data from Twitter, Discord, and Telegram in the next version.