Comment on page
How does it work?
How is the Cred Score calculated? What factors influence the score? Do some factors matter more than others?
Our machine learning credit score model is comprised of a set of features that have been selected to capture the holistic view of an account's on-chain risk. When combined, they create a predictive score that quantifies the creditworthiness of an account.
As the web3 ecosystem continues to grow and mature, we are always looking to add features that inform the model and improve the score. This process of continuous development is driven by our search for new data sources as they emerge, including new lending protocols, chains, and broader web3 projects.
We believe that central to the power of the Cred Score is ensuring that our partners, and beneficial owners of the accounts we score, understand the features that inform our model. This allows for projects to best use the score to quantify risk, and empowers individuals to take action to improve their score over time.
Revealing the individual features that contribute to the score would potentially lead to people being able to 'game' the score, and improve their score without improving their underlying fundamental behaviours. In order to preserve the score's integrity but also provide transparency, the features are grouped into high-level 'Credit Factors'.
A table of the Credit Factors, alongside a more detailed breakdown of each factor is listed below:
These features relate to an account's previous borrowing and repayment behaviour. These features span a variety of lending protocols and take into consideration not only successful repayments and historical liquidations, but also where relevant, other examples of 'adverse events' including previous examples of late payment and delinquency.
Whilst liquidations are penalised, they are also taken within the context of the account's entire borrowing history, such that an account with two liquidations for every ten borrows will score more highly than an account with two liquidations for every hundred borrows. Patterns of borrowing and liquidations are considered also, for example two liquidations followed by ten successful borrows is considered differently from ten successful borrows followed by two liquidations.
If an account has borrowed across multiple lending protocols, then the features within 'Borrowing History' capture the entire scope of lending history. If an account has had a liquidation on Aave, but multiple successful repayments on Compound then it would still be possible for the account to score well in the 'Borrowing History' factor.
Given that these features directly correlate with likely future repayment behaviour, this Credit Factor has the highest weighting in our model. All other Credit Factors being equal, an account that has exhibited good prior repayment performance will score more highly than an account that has never borrowed before.
Account Composition features relate to the assets and tokens that are held within the account, it gives valuable insights into the behaviour and risk profile of a user, and so this group of features carries a high weight in the model.
These features group tokens into categories that capture the reputation, utility, and volatility of a token and the implied risk of holding them. The 'Account Composition' factor seeks to illuminate the different risk profiles associated with how an account is curated, understanding that there is a difference between an account that holds a high number or stablecoin tokens compared to one that holds a high number of yield farming tokens.
These features relate to the value and volume of transactions associated with an account, as well as the account's overall balance. We look at the value and volume of transactions in and out of a account to understand the frequency and scale of an account's involvement in the ecosystem.
Whilst a high balance and a large number of transactions do not necessarily indicate strong repayment behaviour for an account with no borrowing history, all other factors being equal, an account that has a high balance and lots of transactions will score higher than an account with a low balance and fewer transactions.
Our model applies a medium weight to this factor, acknowledging that whilst important this factor does not explicitly relate to prior lending history.
These features are linked to an account's involvement in different type of projects within the web3 ecosystem. The Interactions Credit Factor primarily examines the NFTs within an account, gleaning insights from the type, value and number of the NFTs to understand the communities that an account is a part of, or associated with, and the risk insights that can be taken from that.
The Interactions Credit Factor has a medium weight in our model given the prevalence of NFTs within accounts, and the strong signals that they can provide about the way that the beneficial account owner interacts with the ecosystem.
The features within the Trust Credit Factor seek to capture the web3 reputation of the account, by assessing whether the account likely belongs to an institution, human, or bot, and whether the account has any identity attestation NFTs or has interacted with protocols that require KYC.
Trust and transparency are crucial to creating a healthy ecosystem, and all other credit factors being equal, an account that has demonstrated a strong reputation with a human owner will score more highly than an account operated by a bot.
However given that we acknowledge the entire range of reputation and identity within the ecosystem from pseudonymity to fully KYCd, this Credit Factor has a low weight in our model, to ensure that pseudonymous accounts with strong performance in other Credit Factors can still score well.
This factor groups features related to the most recent lending behaviour of an account, and the trends associated with it. The 'New Credit' Credit Factor provides a window into the 'live' creditworthiness of an account, acknowledging that an account's borrowing activity can change over time either as the user's behaviour and motivations develop, or as the macro-environment within crypto changes.
Given the nascent nature of the lending ecosystem within web3, and the fact that new accounts are constantly interacting with lending protocols for the first time, this factor currently has a low weight in the model, but we anticipate this will increase over time as more data and trends become available.