The unsecured consumer lending business is interesting. It started off being relationship-driven, community-bank based. Perhaps this lender was a bank, perhaps they were not, but it all boiled down to a direct personal connection with the borrower. Sure, there were some sketchy lenders along the way, but that’s par for the course. At some point, unsecured consumer lending shifted to an online application where the borrower’s data is compared against a massive database of other borrowers’ data to determine whether the lender wants to lend to this borrower, and at what price.
In the 2000s, peer-to-peer lending was a thing — a company would do the underwriting work, but the capital used to fund new loans would be from another person. So if you were to apply for a loan from this company, this company would underwrite you based on your personal characteristics and other statistics (like loan request size, time of day, season of the year, whatever else data they can collect) to determine how likely it was that you’d pay back the loan (or inversely, how likely it was that you’d fail to pay back the loan), and would set the interest rate accordingly. You’d get the funds for the loan you took out, and eventually would have to pay that back with interest. However, in a P2P model, your experience may remain unchanged, but the lending company would source those funds from another person, willing to invest their capital given the interest rate.
The main issues with the P2P approach was that retail capital was extremely difficult to predict, which made it equally as hard to scale. So, P2P lenders started turning to institutional sources of capital, via asset based warehouse facilities, flow purchase agreements, and term securitizations. Today (as of early 2022), these P2P lending industry has been called the marketplace lending industry (MPL), and nearly 100% of funding capital is institutional.
So back to the topic at hand — how do we make this Web3?
DAOification of the Underwriter
To start — perhaps we DAOify the underwriting function by making the underwriting function a competition amongst community members.
There will be a virtual sandbox with a database of historical consumer loan and performance data, all anonymized. DAO participants, either individually or in teams, will use a subset of that data (perhaps 60–70% of it) and will construct algorithmic underwriting models (could be a static, statistics based factor model, or a machine learning based model, dealer’s choice) that run within the confines of the virtual sandbox. Each of these models will compete against each other based on several criteria, which could include:
- overall expected loan ROA
- volatility of expected loan ROA
- time delay from data retrieval to pricing
- memory resources required to run the model (i.e. scalability)
- and possibly other criteria
The models will be initially run using the remaining 30–40% of the data in the virtual sandbox to determine the top 1000 underwriting models.
For the next 3 months, these 1000 underwriting models will be tested against the existing underwriting model on new originations, as a check to see whether early performance on new originations using these models would deviate from the expected performance projections based on historical data. This would narrow the underwriting models to the top 100 (perhaps the test here is deviation from projections, as well as the ROA and ROA volatility).
From months 4 to 9, the top 100 models will run together on new originations going forward, with an aggregator that starts with equal weighting across all 100 models — perhaps the way this will work is some sort of majority consensus in term, pricing, and loan size within a certain bandwidth (i.e. it needs 75% of the models to be within a certain bandwidth of each other for it to be applied). This model will be applied to a small portion of the production (perhaps 10%) and compared to the existing model.
At the end of month 9, the bottom 25 models that were outside of the majority for most of the underwriting instances will be banded together, and the bottom 25 and top 75 bands will be backtested to determine default prediction probability.
- If the bottom 25 beats out the top 75, then all 100 models will be retained.
- If the bottom 25 does not beat the top 75, then the bottom 25 will be dropped.
From months 10 to 12, the platform’s loan originations will continue with either 100 or 75 of the models being applied to 10% of the production.
At the end of month 12, the top 50 models that were within the majority for most of the underwriting instances will be banded together, and the bottom 25 or 50 models will be backtested to determine default prediction probability.
- If neither band beats the existing model, then all 100 models will be discarded and the existing model will be kept.
- If the top band beats the bottom band and the existing model, then the top band will be kept.
- If the bottom band beats the top band and the existing model, then the bottom band will be kept.
Going forward, the winners of the underwriting competition will receive their prorated portion of the percentage of the servicing fee.
