Aave Account Protection Score
Deep Dive into Account Risk Assessment
Introduction
We recently received a grant from Aave Grants DAO for building the Aave Analytics Dashboard. While our aim was to make the dashboard metrics simple to understand, we also categorized accounts based on Account protection score. This needs a deep dive to better understand how it works. It represents a risk score on the account level, based on heuristics that we observed in the empirical data as indicative of an account either getting liquidated or not during major market shocks.
If you’re only interested in a summary of insights we recommend skipping to the Conclusions section.
Methodology
We decided to build a heuristic model while leaving the potential of expansion into a more sophisticated implementation, similar to what we did at MakerDAO on Vault Liquidation ML Model.
While heuristics lack certain properties that a machine learning model has, they can also provide good performance with limited complexity. As per #1 of Google’s Rules of Machine Learning: Don’t be afraid to launch a product without machine learning.
We split Aave accounts on Ethereum Mainnet into three categories: Low risk, Medium risk and High risk. We also created a similar heuristic model at MakerDAO which we later validated empirically as being a good measure of vaults’ likelihood of being liquidated.
For Aave accounts, we look at multiple facets of an account position, both current state and historical behavior:
- price drop percentage in volatile assets an account could experience without any proactive health factor increase — liquidatable price drop percentage
- account’s historical protections. An individual protection is defined as each occurrence when the account owner prevented a liquidation by increasing their account’s health factor (deposit, repay, etc.) — number of historical liquidation protections
- number of transactions in the last 7 days — number of 7 day transactions
- for pure stETH recursive positions (only stETH in supply and ETH in borrow) we compute the additional stETH:ETH depeg percentage that an account could experience without any proactive increase in its health factor — additional stETH discount percentage buffer
- number of historical liquidations
Low risk accounts are defined as:
- (liquidatable price drop percentage >= 50%
- OR number of historical liquidation protections >= 5
- OR number of 7 day transactions >= 10)
- AND no historical liquidations
If the conditions for a Low risk account are not met, the account can then either be scored as Medium risk or High risk.
Medium risk accounts are defined as:
- (liquidatable price drop percentage >= 30%
- OR number of historical liquidation protections >= 3
- OR number of 7 day transactions >= 5)
- AND no historical liquidations
If the conditions for neither a Low risk account nor a Medium risk account are met, the account is scored as High risk account.
Analysis
At a high level, we can look at supply distribution across different protection scores. Keep in mind that currently, $2.2 billion (33% of total) is supplied without any borrowing. All of these accounts are scored as Low risk.
Around $6.4 billion of account supply is scored as Low risk. These are usually accounts with high health factors or are highly active in their position management. While Medium risk and High risk accounts’ supply is only 6% of all supply, it totals to almost $400 million of collateral value.
Beyond total supply, we can also dive into how increasing price drop percentages could cause increasing amounts of collateral to be liquidated. The chart below can be found in our dashboard.
To create insights that can approximate actual protocol mechanics, we incrementally decrease prices of both supplied and borrowed volatile assets. If the health rate falls below 1 we compute liquidatable amounts from the current account state. We either take 50% of the individual account’s largest borrowed asset or $5 million as the liquidatable amount, whichever is lower. We chose $5M as the cap because historically, large accounts ($100 million+ in supply) experienced liquidations in multiple smaller chunks which didn’t exceed this amount (or if it did, it was negligibly higher). When the liquidated amount is applied, the account’s health factor rises above 1. In case it falls below 1 again because of further price decreases, this triggers another cycle of liquidation logic explained above. There are additional nuances about our modeled liquidation process that are beyond the scope of this post.
In the chart below, we can see that beyond the 50% price drop mark, collateral tends to get safe rapidly. If we focus on the price drop below 50% we can see that most of the collateral amount at risk is scored as either Medium risk or High risk. Fortunately, this amount is split across many different accounts which can decrease protocol risk significantly (less likely for all of these accounts to be unresponsive in case of an actual price drop).
When looking at the largest accounts scored as High risk, we can see that there are often two main reasons for the score: either low liquidatable price drop percentage or historical liquidations.
It’s interesting to see that some accounts have both a high number of protections and also liquidations. This could be useful feedback for the existing model improvements where we would be validating on recent history whether account owners improved their skill of position protection over time. There are also other potential reasons for both liquidations and protections such as manual position management which would cause the account to be more likely to be liquidated in a specific time zone. For this reason it’s crucial to validate model logic with empirical protocol behavior to continue improving its performance of risk assessment.
Currently, pure stETH/ETH leveraged long (only stETH in supply, only ETH in borrow) positions contribute to 20% ($310 million) of total stETH supply. This percentage used to be higher but due to good risk management awareness in the Aave community, this decreased in the last weeks. Meanwhile, these positions are theoretically most at risk of liquidation in case of substantial additional stETH:ETH depeg.
We can see below that most of the total supply has a buffer of at least 10% in additional depeg percentage.
Beyond this stETH analysis, we also created a stETH at risk chart (available on stETH price drop (depeg) tab). It simulates seized amounts of stETH with additional depeg percentage, similarly to the above presented liquidation logic. The difference is that only stETH’s price is incrementally decreased, not the prices of all volatile assets in the account position.
A unique signal in our model is looking at the account’s historical protections against liquidations. So far, 410 accounts protected themselves successfully. That’s less than 2% compared to all accounts that currently have some borrowed amount (~25 thousand). This might seem low at first but given that most accounts keep a high health factor, haven’t experienced a major market shock, and if they did they were likely to be liquidated, this is not surprising. Also, our definition narrows down the number of eligible events significantly to minimize false positives.
We define a protection by taking each historical account event that increased the account protection score and then look within a specific window whether the account would’ve been liquidated if there was no counter-action to the decrease in account’s health factor.
The chart below shows a histogram of the number of accounts per number of protections. Most (⅔) of accounts protected themselves once while 30% protected themselves between 2 times and 5 times. Somewhat surprising is that more of these accounts are not scored as Low risk. That’s caused by almost 40% of the sample accounts that protected themselves also experienced liquidations in the past. This means that these are a subset of accounts with high risk tolerance. They most often keep low health factors to maximize capital efficiency, sometimes at the expense of liquidation events. For this reason we also increased the threshold at which an account is deemed to be well versed in position protection (5 instances for Low risk, 3 instances for Medium risk, as per above described methodology).
Conclusions
In this post we introduced Aave account protection score methodology and shed some insight that can be extracted by surfacing valuable risk signals. An important property of this heuristic model is that it provides individual account risk assessment. Methodology can further be expanded into a more sophisticated model.
We showed that most of the current supply is safe from liquidation risk. Meanwhile, when looking at collateral at risk below 50% price drop level, most of the collateral is scored as either Medium risk or High risk. But considering that this is split across many accounts this does decrease protocol risk of bad debt. Besides that, the liquidatable amounts are too small to pose significant systemic risk by triggering cascading liquidations. Additionally, stETH positions are safer compared to recent history due to good Aave community awareness of best risk management practices, especially below 15% additional depeg percentage.
While most of our work focuses on studying risk at MakerDAO, the DeFi ecosystem continues to be increasingly interconnected with many sources of potential systemic risk. This is especially clear after the D3M implementation through which the Maker system inherits part of Aave protocol’s attack surface. Currently, there is no DAI exposure from Maker’s side because of low Aave DAI supply rate and MakerDAO’s decision to be risk off before the Ethereum merge event. This is likely to change soon.
Through publicly shared risk analysis of different protocols we can make better, more informed decisions throughout the ecosystem.
Acknowledgements
Thanks to Primož Kordež and Marko Štemberger for reading earlier drafts of this article and providing valuable feedback.
