Credit Scoring in the Cryptocurrency Ecosystem

Proponents of cryptocurrency often tout it as ‘banking the unbanked’. While there are many possible advantages of providing increased access to banking functions to people across the world, cryptocurrency only provides one:  the ability to transmit money quickly and (sometimes) cheaply.

To truly provide banking services a system would need to extend to borrowing and lending. Traditional financial institutions have been evolving the way money is borrowed for centuries, so why not do the same thing with cryptocurrency? A primary tool that banks use when lending is the credit score.

Let’s dive into what a credit score is, and whether it could be useful in the cryptocurrency space.

example scorecard

Credit scoring models have long been vital to the banking industry. They are highly interpretable and easy to deploy and are competitive against other models in their predictive power. Credit scoring models have been used since at least the 1850’s.

Scorecards determine how much credit risk a potential borrower carries, and they’ve seen mass adoption because they:

As consumers we see our credit score as a number between 300 and 850, where a higher score is ‘less risky’. These scores come from models that use data about an individual to assess their ability and likelihood of repaying borrowed money. Different banks and credit rating agencies employ different models, but the models rely on similar features that include:

While there may be alternative data that lenders can use to assess credit, there are also severe restrictions on what data they can’t use in order to promote equitable access to credit. All of this data, though, is associated directly with the individual person who is seeking credit.

The data from a great number of people and their repayment history is used to build an underlying model (like a logistic regression) to classify the likelihood of a new person defaulting on a loan, and then that model is scaled to provide a score. Regulators require that the impact that each feature has on the overall credit score be transparent and interpretable, so it is clear what an applicant could do to improve his or her credit score.

Currently, to obtain a loan, a borrower applies, and if approved is offered an interest rate and loan terms. The applicant must provide their personal details so that their creditworthiness can be assessed  (likely by a scorecard) by a model. But what if this could be done without revealing your personal details?

One allure of cryptocurrency is its anonymity.

A user’s true identity does not ever need to be verified. This is true for transacting directly with cryptocurrency, and it is also the case with most lending and borrowing platforms that use cryptocurrency.

How, then, do lenders assess the risk of the borrowers? In most cases today, the lenders don’t assess that risk, they simply force the borrowers to offer collateral for the loan. If the loan goes into default, then the borrow loses the collateral. Much like a traditional mortgage uses a property as collateral, in this decentralized finance (defi) space, the collateral for borrowing one cryptocurrency is typically a different cryptocurrency.

When it comes to decentralized finance a lender may not be able to see an applicant’s credit history or say, marriage status, but the ledger that records most cryptocurrency transactions is completely open. This means that crypto transaction history is available for anyone to see and use — a trove of useful data. Just as credit scoring models are trained using historical data about the applicant, could this open ledger of all transaction history be used to assess an applicant’s credit worthiness?

What is Data Wrangling and Why Does it Take So Long?

For each “wallet” (individual), features can be extracted using the transaction history (like here) and then used to model whether a user will repay a loan or not. There are many different cryptocurrencies, and there are many protocols that have been written to facilitate lending and borrowing (most of them are open source). Aave is one of the most popular, and the full history of lending and borrowing is public and open for anyone to analyze.

Since the crypto transaction history and loan status data is available and open, anyone can use it for analysis or model training.

In this very basic example I’ve extracted all lending data on Aave over the four-month period ending April 30, 2022. I’ve aggregated the transaction details from January through March and used that data to train a model that predicts whether a user will default on a loan in April. The summarized transactions are a simple count of each type of transaction a user had during the Jan-Mar period, as well as a count of the unique number of tokens (cryptocurrncies). The ‘target’ is whether the user repaid (0) or defaulted(1) on a loan in April.

This is a comically simple model and is used just to illustrate the type of data that is available and show that it is possible to build a model to assess likelihood of repayment. Although there were many liquidations (or defaults) in this history, recall that all of these loans are fully collateralized, so the lenders had virtually no risk.

The problem with using a credit scoring model like this to offer under-collaterilized loans is that it is trivial to ‘game the model’. While each unique user has a fully available transaction history, that user ID is not associated to any real person or verified identity. A bad actor could quickly and cheaply generate hundreds or thousands of user ids, execute transactions amongst them to increase their credit score, and then borrow money to never repay it. That bad actor would suffer no consequences because the user ids are pseudonymous. They could simply create new user ids and start the process as many times as they like.

If decentralized finance ever has the capability to provide credit to the unbanked, then a reliable credit scoring system is imperative. There is a rich transaction history available, and there is an opportunity for creative data scientists to find ways to model credit risk that aren’t easy for bad actors to game. One promising way forward is to employ graph analysis to uncover likely rings of ‘fake’ transactions. With the open ledger, the transaction history of a user doesn’t just show amounts, but it is easy to see the user on the other side of each transaction.

Of course, borrowers could associate their personal identity and credit history with their cryptocurrency wallet address, but that isn’t a possible (or palatable) solution to many would-be crypto borrowers. We’re not the first to get excited about this challenging problem and vast depot of transactional data; there are many firms that are focused on developing cryptocurrency credit scoring models, to name a few:

LedgerScore

RociFi

Credefi

Quadrata

A winning and useful model could help provide credit to people around the globe that don’t even have access to a bank, so long as they have an internet connection. A credit scoring model will also make the lending markets much more efficient, with less capital tied up idly as collateral in loans to safe borrowers.