Building Credit Metrics in the DeFi Universe: An Experiment

We have witnessed the rapid growth of DeFi developments in 2020 and 2021, especially in the lending and trading sector. The Aave and Compound protocols have been the leading protocols on collateralized decentralized lending and are currently dominant in DeFi. The benefit of DeFi is that all activities are on-chain and everything is fully operated by computer codes, which makes the whole process transparent and avoids risk from intermediaries.

The standard procedure for collateralized lending on the blockchain is to post collateral in the form of crypto to the protocol vault. Then the user can take out a loan in another crypto. Typically the loan value is lower than the deposit value, which provides a safety buffer in the case of volatility. It’s the same idea as getting a collateralized loan from a bank. Recently a few protocols have begun to enter into uncollateralized lending territory, though they are aimed at a much narrower audience, i.e., they work through intermediaries or only lend to reputed borrowers to contain the credit risk.

However, when we look at the original purpose of DeFi – to provide financial services to the millions of people who can’t get access to traditional banking – it seems that we are far from reaching that goal. Lending is one major service of the banking sector. Under the current collateral model, with loan amounts lower than the deposit value, it’s almost impossible for impoverished people and startup companies to get access to the lending part of the DeFi world. That’s where uncollateralized lending can be extremely helpful. Not to mention that credit-based finance is frictionless and is the type of lending we use most frequently in our daily lives – compared to the few transactions we might make using mortgages once every few years, to purchase an apartment or car, we use credit cards way more often, mostly on a daily basis. Uncollateralized credit-based finance is the lubricant of our lives.

The key problem with DeFi uncollateralized lending is the lack of a credit system in the DeFi universe. Unlike the financial world we interact with today, DeFi is anonymous in nature, which makes it hard to track down the individuals and corporations behind each wallet address. This significantly raises the risk of moral hazard. There currently is no regulation in DeFi, which makes it extremely difficult to seek legal actions against delinquency.

The key, and probably the largest, missing part is a credit metrics system for each account, which could screen risk from the beginning. Without this, we are almost blind when dealing with loan requests. In the real world, people have credit scores, and such credit scores are the key input for lending, credit cards, and mortgages.

A credit score is based on each individual’s credit report, which contains data on credit card transactions, income, etc. The FICO score, for example, utilizes a model that analyzes the data from credit reports and generates credit scores that are most widely used by lenders. In addition to credit report data, a few recent developments in emerging markets have shown that alternative data can improve credit score accuracy in an environment where traditional credit scores are not available.

Zhima Credit, introduced by the Ant Group – a subsidiary of Alibaba – has been said to utilize data from online shopping, utility payment, and social relationship to create its own credit score system, and the result has been promising. According to a research paper, the Zhima Credit score could indeed improve classification performance within an experiment involving over 20,000 online loans. However, there have been intensive discussions on privacy protection issues related to this type of credit score.

There are a lot of similarities between the DeFi world and the market Zhima Credit faced when it first began to introduce its credit score system in China, where there was not a lot of banking data available. Comparatively, DeFi has the advantage of an open data source. All data is available on the public ledger, and thus it provides the best data available out there. It provides the perfect ground for big data analytics and machine learning. With wallet owners’ consent, their credit score could be readily available, as long as we have the tools and models ready to use. Given DeFi is the foundation of the metaverse, a credit system based on DeFi could support future activities in the metaverse as a whole.

The first step is to utilize the Decentralized Identity (DID) system to link an on-chain address to an identity, connecting real people with online accounts. Once the link is confirmed, the off-chain identity and related creditworthiness data could be brought into the on-chain world. For those who have credit scores (a FICO score, for example), this would create immediate creditworthiness for their DeFi user profile. However, most users do not have such a credit report, and privacy can be a concern. Thus, an on-chain credit system is necessary to fill in those blanks. It’s worth noting that, even in those scenarios, the DID system still can play a very important role in connecting a wallet address to a real person and reducing moral hazard and fraud risk.

The next step is, how can we build a framework? In the case of the FICO model, its data sources can be grouped into five categories: payment history (35%), amounts owed (30%), length of credit history (15%), new credit (10%), and credit mix (10%). To start with, it’s possible to create a very rough but similar model in DeFi:

• payment history – this could be represented by on-chain transactions or NFT purchase history;
• amounts owed – there is no good proxy here; however, we could try using the total debt exposure on chain;
• length of credit history – the length of time that the user has ever used a loan;
• new credit – new credit on chain;
• credit mix – this could be anything that the user has outside the above four metrics.

It’s apparent that this is not a suitable model to fit on-chain activities well. However, we could start from here and develop models that give us certain expectations about default probability. Then we could revise our model with data we observe.

A better, and the ultimate, solution would be to create such an uncollateralized credit lending market in DeFi. In this way, we could observe, collect, and process the default output data along with data related to on-chain activities. We could then train the model utilizing supervised machine learning methodology: We could maps inputs (on-chain activities data + available off-chain data) to an output (default probability based on example input-output pairs).

Carrying out a full-scale lending business brings significant risk; all money can be lost during the process. Thus, it might make sense to start with a small group experiment. The experiment could host a small group, for example, 1,000 people and last for six months. Participants could have a credit line of 1,000 USDC upon their request and standard terms for a credit line would be applied. We could observe the participants’ repayment activities to collect data in the same way as credit agencies collect data in real life. With such data, we could create the first model for DeFi credit scores. This then could lead to a second experiment with more participants, higher credit limits, and more tailored lines based on the score system. After several such experiments, we could expand such a business model to a wider-reaching population with a high confidence level. When issues are found during each round of the experiment, the model could be revised and updated for the next round.

This project could run under a corporation in the beginning, and it could later be transformed into a DAO (Decentralized Autonomous Organization). The algorithm used to calculate such a credit score could be kept undisclosed, as scammers can easily manipulate a system once it’s public. The initial funding for the experiment could come from a group of investors or an asset pool created by crowdfunding. This could help to alleviate the risk for each participant. Although a lot of things could go wrong with this experiment (for example, the data quality could be problematic, the risk control model could be messy, and the data-training procedure could be erroneous), the benefit of such an experiment could be enormous for the whole DeFi world.

A reward system would also be an important factor in DeFi lending. A good credit score could be rewarded with a specific NFT, or could lead to a whitelist for NFT minting. People with good DeFi credit could borrow money or NFTs for a few days using a simplified process. A high DeFi credit score could also be rewarded by entry tickets to metaverse events. This would be similar to the way a high credit score can function positively when renting an apartment or applying for a new mortgage.

Overall, credit is an asset, and it can generate positive returns. It’s still missing from the DeFi universe, even after the total DeFi market has grown to 100bn USD. It’s time for us to work toward it.

Bi @trtwarrior, former derivatives trader on Wall Street and director of a regulated crypto lending firm.