RociFi NFCS For Capital Efficiency — Radiant Capital

2022-12-06

Dec 6

By using RociFi NFCS, Radiant Capital can increase revenue by 5.19% without increasing liquidation risk.

RociFi’s NFCS seeks to rank potential borrowers based on the likelihood of repayment if they were extended a loan. Using a plethora of historical data collected from major DeFi protocols from 2018–01–01 up to 2022–10–31, we have included Radiant data** into these transactions, and have been able to refit and estimate the risk of default for borrowers extended over-collateralized loans.

Using 2022–08–05 to 2022–10–31 as an evaluation period, we define “good” accounts as those that did not default on their loans due in this period, and “bad” accounts as those that did. To better understand if more creditworthy borrowers are less likely to get liquidated, we can look at the distribution “Actual Goods” vs. Actual Bads” for the broader ecosystem first. For this analysis we exclude any borrowers who have only ever taken 1 DeFi loan, regardless of whether they have paid it back or not[1].

On the broader distribution of protocols covered, roughly 71% bad borrowers[2] are ranked as 10. ~96% of bad borrowers get ranked a 6 or higher. Although 26% of good borrowers[3] ending up on score 10, we can see there is a gradual increase in the ratio of good to bad borrowers as the score approaches 1.

To get a better understanding of the impact specifically to Radiant we can take a look at the distribution of borrowers who are specifically using radiant protocol. Again, any borrowers who have only ever taken 1 loan are excluded from this analysis.

For Radiant, the distribution of borrowers tends to lean more to the right. Roughly 81% of “bad borrowers” get score 10 with 65% of borrowers who did not default also ending up on score 10. About 95% of “bad borrowers” end up on scores 7–10. Part of the reason for 65% of “good borrowers’’ ending up on score 10 has to do with both length and depth of credit history. 50% of these borrowers had fewer than 5 loans and all of them had accounts less than a week old at the time of scoring. We expect as Radiant Capital matures more borrowers would naturally move down the curve.

The bad loan rates for the broader ecosystem and Radiant are as follows

On the broader ecosystem, the risk of liquidation remains well below 1% up to score 6, however for Radiant borrowers this risk is generally much higher, with the risk already at 1.05% on score 2. On both we can see that the risk is increasing with score, with the exception of some confusion around score 3 & 4, 6 & 7 on Radiant[7].

As we can see, utilizing the RociFi NFCS would allow Radiant Capital to reduce the risk of default at higher LTVs by limiting those LTV’s to borrows with more creditworthy scores.

In order to estimate how much capital efficiency would be improved by, we look at the history of radiant lending transactions from the Radiant Subgraph, and run revenue simulations based on the lending transaction history derived from the subgraph.

The ROCIFI NFCS model was retrained and refitted using data collected from all covered protocols and in addition we have Radiant. An LTV curve by score is proposed below based on an assumption around liquidations slippage, implied default rates, and asset standard deviations. Furthermore, the LTVs are intentionally aggressive to generate the most aggressive liquidation scenarios, thus most conservative projections.

For stable collateral assets (USDC / USDT)

For volatile collateral assets

The formula for picking the LTV curve here was as follows

The idea here being assuming that we don’t want LTV’s to exceed 100%, using a 5% margin of error, we scale back the LTV by the Expected Default Rate, and 2x the daily standard deviation of price returns. The scaling factor was selected in each instance so that whenever volatile collateral is involved, in the worst case we assign an LTV of 60%, this is consistent with the current maximum LTV on Radiant. Given default risks and price volatility, we believe this LTV curve conservatively gives sufficient runway for fast price movements that may cause problems for liquidators in light of default risks.

Interest Rates

Following these LTV’s we simulate lending revenue’s using approximate interest rates based on what was observed on Radiant, those APRs used for simulation are as follows

Using the interest’s rates above and proposed LTVs we align them onto a dataset of Radiant’s borrowers by score, and estimate the hypothetical revenue(s) that could be generated by utilizing the proposed LTV’s.

The simulation procedure is performed as following

Isolate all borrow, repay and liquidate transactions
For transaction, align the appropriate score for that address
For each transaction, using the score computed for that address, align the proposed LTV based on score and token
For each transaction, using the token, align the interest rate[9]
Using the ratio of the proposed LTV and Artemis’s LTV, lever repay and borrow transaction sizes by this ratio to compute the theoretical increase in size that would be produced by the affects of leverage[10]
For each address, by each token, by each date, compute the net outstanding loan owed both using the original LTV and proposed LTV’s
Group these by address, token, and date, and compute cumulative sum of the outstanding loan by day
Estimate the daily revenue by scaling each day’s outstanding amount by the daily interest rate for that token. If the net outstanding on a particular day is negative, we treat this whole quantity as interest income
Compute the incremental revenue as the difference between the original revenue estimate and the revenue estimate produced using the enhanced LTVs

Results

We estimate if Radiant had used RociFi’s LTVs as proposed, it could have generated $12,117,061.89 in revenue[11].

For comparison, we also look at the revenue we estimate Radiant would generate using current base LTV’s and the interest rates proposed in this document. Since July, we estimate this would have been $11,560,362.28, annualized.

Obviously, increasing the leverage borrowers are allowed to take enhances the potential return given the same interest rate. The increase in revenue is about 5.19% increase, which represents the effective increase in capital efficiency to the Radiant ecosystem.

We can take the difference in these quantities, and look at the incremental revenue by month

Overall, we estimate that using the RociFi NFCS, Radiant could have generated an additional $556,699.61 in revenue.

Loan Distributions

Currently, we rank 89% of Radiant borrowers as 10. The risk observed among Radiant Borrowers is generally much higher than the broader ecosystem. Generally speaking the feature values observed for Radiant borrowers do indicate they are of higher risk than the average borrower that uses AAVE / Compound. This is not to say all Radiant borrowers are risky, but a sizable minority are. Regardless, the RociFi NFCS allows for credit migration based on user behavior, both good and bad. Those Radiant borrowers who use the protocol responsibly, along with other protocols, will see their scores will increase over time.

It is worth noting that although most Radiant borrowers score closer to a 10, the majority of borrowing volume comes from less risky better ranked borrowers. 47% of borrowing volume is coming from borrowers ranked a 4 or better.

Using the proposed LTV curve principal losses were minimal. We observed only 1 instance of principal loss estimated at $2,917.49. Under the simulation, this occurs when there is an outstanding balance due on the loan, but insufficient collateral to pay off the outstanding balance at the time of liquidation.

There were some instances where the loan was liquidated successfully, but the protocol losses interest payable as the loan is no longer active. This should not be confused with true losses as the lender has gotten back all the principal they lent out.

Hypothetical example, 20% APR is owed on $100 and $110 in Collateral is posted. The loan experiences liquidation and accounting for repays and liquidation proceeds, $102 is recovered at time of liquidation. If $120 would be owed assuming the borrower held the loan for one year, there is $18 in un-accrued interest although all $100 of principal is recovered.

These events reduce the amount of potential revenues earned by lenders. Below we show this by the month.

Next, we look at the unearned potential interest revenue by token

The vast majority of unearned interest revenues are coming from WETH and WBTC. Although potential principal losses are minimal, we can also consider a more conservative LTV curve to compensate for elevated levels of unearned interest revenues.

As a percentage of of borrowing volume we get a similar story

As a percentage of borrowing volume, unearned interest revenues are highest on WETH, followed by WBTC and USDC. The higher percentages on USDC and USDT relative to dollar amounts of unearned revenues suggests that as borrowing volume increases on USD we could expect more unearned interest revenues.

However we should bear in mind that the evaluation period for these lending volumes encompasses a bear market, where generally lending volumes for USD linked stable coins would be lower. During a bull market rising values of risk tokens would likely drive increased demand to borrow stable coins, and would also likely encourage repayment to avoid liquidation penalties assessed on that same rising collateral. Thus during a bull market we’d expect unearned interest revenues as a percentage of borrowing on stable coins to be lower.

Key Assumptions

Increases in leverage offered to borrower does not affect the probability of default as long as the LTV does exceed the proposed LTV by score
For loans that get liquidated, if there are any repayments before the liquidation occurs, we assume that those repayments will also get scaled up by the leverage ratio
All loans are originated at exactly the proposed LTVs
Constant scoring of borrowers over the lifetime of the simulation
Constant rates over the lifetime of the simulation
Interest revenue can be collected daily against outstanding loans via collateral offset

For assumptions 1–4, we anticipate that actual revenue increases could be higher than what is displayed here. As stated prior, we’ve already spotted areas in our analysis which can be optimized for better risk-adjusted gains.

Incremental Revenue gains are limited due to generally higher levels of default observed among Radiant borrowers of all scores. We suspect as Radiant continues to grow, the risk of Radiant borrowers would converge towards the general risk of larger protocols such as AAVE and Compound.

Overall by using RociFi NFCS, Radiant will be able to increase capital efficiency available to its borrowers, which we believe will grow the size of Radiant Protocol and value proposition to stakeholders and the community.

[1] Under the current release of RociFi NFCS, borrowers who have only ever taken 1 loan ever, regardless of size, are assigned a score of 10.

[2] A “bad borrower” here is defined as any borrower who experienced a liquidation in any amount over a 2 month period prior to the end of the evaluation period.

[3] A “good borrower” here is defined as any borrower who has not experienced a liquidation in any amount over a 2 month period prior to the end of the evaluation period.

[4] Although the risk on score 1 was estimated at 0%, there were only 22 observations for radiant borrowers. Given Radiant borrowers tend to be riskier than the broad ecosystem, we’d expect the risk on score 1 to be higher than the broader ecosystem on score 1

[5] Excludes borrowers with only 1 borrow on record. If these borrowers were to get included, the risk on the broader ecosystem on score 10 would be ~2.41%

[6] Excludes borrowers with only 1 borrow on record. If these borrowers were to get included, the risk on Radiant for score 10 would be ~6.25%

[7] There were 252 observations on score 3 and 236 observations on score 4. On score 3, 4 bad loans were observed, and on score 4, 3 bad loans were observed. This decrease in reported risk is likely due to low sample size, and the risk would likely come in higher given a large sample. On score 7 only 9 bad observations were observed on score 7 out of 371 observations, and on score 6 only 12 bad observations were made out of 278 observations.

[8] DAI has a lower maximum LTV than USDC or USDT at 75% per Radiant docs

[9] Note we use the same interest rate for each time period, rather than the historical rate at each point in time. The interest rates used were captured off the front end of Radiant Capital.

[10] Current framework levers levering both borrow and repay transactions for loans that did experience liquidations which could prove incorrect in real-world scenario

[11] This assumes all loans originated at exactly the proposed LTV’s