Aave Protocol Liquidation Retrospective: May 2021
The last two weeks were a significant stress test on the Ethereum decentralized finance (DeFi) ecosystem, which was marked with high volatility, high gas prices, and liquidity concerns. This environment provided an insightful stress test for various parts of DeFi. For example, network congestion was worse than Black Thursday by many metrics. In addition, Compound and Aave held over 100x the total assets since Black Thursday, which could have led to dangerous liquidation spirals. This post will dissect some of the high-level statistics from the Aave protocol as well as future steps. Overall, on-chain borrowing protocols fared well in these extreme conditions!
Centralized Exchange Liquidations
For perspective on everything that happened in DeFi, let’s take a look at the liquidations on centralized exchanges first. The worst day for centralized exchange liquidations in this recent market sell-off was May 19th. In Figure 2, you can see that there was a total of 7.09B worth of USD (2.95B USD happened on Huobi) of long liquidations on centralized exchanges. This number is higher than the value of liquidations that happened on April 22nd but is surpassed by the liquidations on April 17th when over 9B USD was liquidated. Around 4.7B USD worth of positions were liquidated on Binance on April 17th, which could partially explain the drastically lower liquidations on May 19th of 870M USD. In addition to the liquidations, several centralized exchanges also suffered technical difficulties.
On-chain Protocol Liquidations
For this analysis, there will be a focus on May 19th. Although it was not the local minima for the price of ETH, it saw the most drastic intraday move. The peak to trough move of ETH was approximately 41%, from $3400 to $2000. (on centralized exchanges, there were prints of as low as $1700). As a result, both Aave and Compound experienced the largest single-day liquidation total in history.
From May 17th to May 23rd, the total collateral liquidated on Compound was around 260M USD and on Aave was around 310M USD. At that time there was 10B USD and 13B USD worth of collateral on Compound and Aave respectively. Note, the liquidated collateral value has the price marked against the closing date of the liquidation and there could be discrepancies if marked against the nearest price to liquidation given the extreme price movements. The largest single-day liquidations happened on May 19th for both Compound (~$160M) and Aave (~$170M). Below we take a closer look at the specific collateral and repaid assets from Aave liquidations.
In the figure above, the total value liquidated is broken down by the repay. As expected, the majority of the liquidations are non-stable crypto asset collateral against stablecoin borrow. It is also important to scale the total liquidated volume against the total supply of each asset. The scaled liquidation volume ranking is:
- YFI (5.6%)
- WBTC (3.2%)
- LINK (2.4%)
- WETH (2.3%)
In our Market Risk Assessment for the Aave protocol (see page 18, Figure 2b), we discuss the assets which have the most systemic risk given the current risk parameters. The top three assets by percent liquidation were all part of the assets that where our models observed risk. Our simulations with the current protocol settings did predict that these assets could be at risk for large liquidations and consequently for insolvency.
Gas Spike and Network Congestion
From May 17th to May 23rd, the top 50% of gas bids were above 100 gwei and the top 1% of gas bids were consistently around 1000–4000 gwei. Compared to Black Thursday gas prices, on March 11th and March 12th, 2020, the top 50% of gas bids were above 60 gwei and top 1% of gas bids were above 400 gwei. High gas prices will have negative impacts on the Aave protocol (more details in our report). In general, transactions that top-up accounts close to liquidation and liquidation transactions are both incredibly important to the solvency of the protocol. During times of high gas and network congestion, these transactions might not process in time leading to insolvent debt. To assess the impact of network congestion Gauntlet models liquidator and borrower behavior. Notice that in Figure 4, the majority of Aave liquidations on May 19th did not need high gas bids. Yet, over 30 liquidations happened at a gas bid over 2000, which is concerning. With liquidators willing to pay such high gas prices, it ultimately leads to more liquidations and worse user experience.
It is also important to compare borrower behavior during this same time period. In Figure 6 and Figure 7, we look at the top-up (or additional deposits) on the Aave protocol. Note that in Figure 6, all gas bids above 2,000 GWEI are included in the 2,000 GWEI bin. Although there is approximately 10x the number of the top-up transactions during the same time period, there are significantly fewer transactions with over 2,000 gas bids.
Figure 8 is a valuable comparison of the total top-up amounts compared to the relative difficulty for each collateral asset. The average gas bid utilized to top-up certain assets (LINK, CRV, YFI, UNI) was significantly higher than others.
There were 18 distinct liquidations on the Aave protocol that preceded failed top-up events. Top-ups are debt repayment or collateral addition attempts ahead of liquidations. This metric definitely does not account for the users that were discouraged from topping up (in terms of deposits or repays) because of extremely high gas prices, but it does shed light on how the network congestion affected users of the protocol.
Also, over half of the failed repays were [Out of gas] failures. Approximately 35M USD worth of collateral was liquidated with failed repays.
Figure 10 is a characterization of some of the liquidators that were able to liquidate accounts with failed top-up attempts. Notice that these liquidators are quite sophisticated in both their diming (incrementing gas bid by small denominations) and private transaction behavior.
Insolvencies
A focus of our Market Risk Assessment was understanding and simulating protocol resiliency. One of the key metrics used to determine resilience was net asset insolvency, or underwater debt.
At the beginning of March, when the deposit size and utilization of Aave V2 was beginning to overtake V1 in TVL, there was already about 14,000 USD worth of insolvent debt on the Aave V2 protocol, but 10,000 of the insolvent debt is one account that borrowed ENJ, UNI, BAT, and ZRX. Most were very small positions that liquidators could not profitably liquidate. As of today, there is approximately 60,000 USD worth of insolvent debt on the Aave protocol. There is only one account with over 3,000 USD worth in insolvencies. The majority of the collaterals used in the accounts were WETH, LINK, or YFI. At a TVL of over $10B, this is less than 0.05 bps increase in protocol insolvency! This is a promising statistic to see for the future of the Aave protocol and it’s users!
Next Steps
By Gauntlet’s main metrics for protocol safety, protocol insolvency and insurance fund slashing, the Aave protocol excelled. In addition, Gauntlet now has a plethora of data to test and validate key assumptions and agent behavior in our simulations. At the same time, the large number of liquidations that occurred are not indicative of an optimal user experience. While Aave has withstood this stress test, these events expose issues the protocol faces:
- How can we improve user experience by minimizing liquidations (total number of liquidations and amount lost to liquidation bonus) in the future?
- What metrics can we continue to monitor that are leading indicators of account solvency apart from collateralization and health factor?
- How did borrower behavior change during the the last few weeks and can the protocol update parameters to promote beneficial user behavior?
One solution to these problems is allowing the protocol to update risk parameters in response to market conditions. In the past, the focus for analysis was solvency and mitigating insurance fund slashing, but that is only one aspect of protocol optimization. As Aave grows and matures balancing safety with maximization of returns for users will be an increasing priority.
(Special thanks to Shaan Varia, Hsien-Tang Kao, and Nick Cannon for their help on this post)
