Loan & Policy Scoring
Evaluating an individual loan and (potential) policy with respect to the collateral type, the internal protocol metrics and the particular features of the originating wallet.
Introduction
We've seen that the Collateral Scoring service produces C-Scores about the collateral tokens.
Now, the Loan & Policy Scoring Service focuses on a particular policy, already existing or a new one (quote), looks at the C-Scores made available by the Collateral Service, looks at the particular features of the originating wallet, looks at the current state of our global portfolio and the incremental impact of the policy instance being examined, and generates L-Scores, measuring the profitability and risk of the policy.
Again, like the Collateral Scoring service, this one is designed to plug-in a diverse set of models both from classic finance and ML world, as long as they output L-Scores when a request is submitted.
The input request will typically include the originating wallet and the position to be protected; while the output can take one of these forms:
Probability of a claim followed by a default.
Probability of a liquidation event.
Price of the policy according to a BSM model or similar theory.
Probability of the request being fraudulent or anomalous, for example if too many similar policy queries were submitted in a shot period of time, we should have L-Scores that prevent approving more of these and trigger an anomaly detection alert.
Any numeric or categorical output is a valid L-Score as long as it can be produced fast enough (in seconds) whenever the service receives a request.
Methodology
Web3 offers a free and rich data field to assess a wallet creditworthiness: previous lending history, token trading features, NFTs and SBTs, etc.
We apply state of the art modelling techniques on this data to build, evaluate, operate and monitor the performance of our credit scoring models.
A champion challenger framework will be adopted where multiple models run in parallel to provide decisioning with predictive power. The latter will generally use a few models as decision points, while the others will generate scores in dry-run as candidates for better performance.
One of the hardest challenges of credit scoring is dealing with selection bias and blind spots. We will address it by combining multiple tactics: approving a small fraction of loans on a random basis; approving a fraction of loans using different orthogonal strategies; using external scores and data from other web3 credit scoring projects as a means to learn about our blind spots; and stress simulations with pessimistic assumptions about the unavailable data.
Example
As an example, the Nimzovitch model uses data from the Ethereum blockchain and predicts if a wallet taking a loan will have a liquidation event in the next 30 days.
As you would expect in a credit model, the most predictive features are the previous liquidations, followed by the number of borrowing events, LTV, and some wallet features like big stable coins transfers.
Note that the features that go into Fico scores are very similar: previous defaults is the strongest, followed by number of open accounts, current credit utilization and age of account.
Models Catalog
Nimzovitch Model
This class of models is trained using XGBoost to predict the probability of a liquidation event in the next n days for a given wallet at the time when they take a new loan.
Model documentation: Nimzovitch 0.0.1
A series of experimental notebooks can be found here
Integration within the Product Ecosystem
Last updated