We posted our methodology and following recommendations here. For V3 Ethereum, we recommend slightly more conservative parameters (LT=96.5% and LTV=93%) and ask for this recommendation to be an option in the coming Snapshot vote.
We would like to highlight a few differences and points of emphasis on the methodology above:
We find that correlation is not an adequate metric to include or exclude assets from E-mode:
For example, two assets might have a high correlation simply because their prices are increasing, but the gap between them might be widening, leading to liquidations. Instead, we consider testing for the mean-reversion of pairwise prices, which eliminates the risk that correlations are spurious. These are sometimes referred to as stationarity tests.
The correlation results presented in figures 4, 5, and 6 may not be very significant due to the 15-minute time frame being too coarse. It may be more appropriate to look at all Chainlink oracle updates, and, if needed, resample data at short intervals (we use 45 seconds).
We find that reported negative correlations (e.g. -0.16 for USDT) are insufficient to discard assets from E-mode.
Could you kindly explain the methodology that Gauntlet follows to set the LTV? We find that the LT-LTV spread acts as a “buffer”, helping prevent retail users from getting liquidated. Although a 0.5% spread is reasonable, USDC/DAI borrowers who took a loan near the LTV could’ve been liquidated three times during the stablecoin depeg events last weekend and four times over the last year. In our methodology, we enforce a sufficiently wide gap, such that borrowers at or below their LTV are not liquidated with deviations of up to $\gamma$ from the mean price.
We have provided our stance on stablecoin E-Mode here.
Before recommending specific configurations and parameter recommendations for the E-Modes on the different V3 deployments, we think the community should align on its approach towards the stablecoin E-Mode in general. We recommend the coming Snapshot have the following options:
Aggressive parameters with a limited number of assets
If the community chooses this approach, the recommendation would be to take USDT and MAI out of the current E-Modes
Conservative parameters and a wider range of assets
If the community chooses this approach, we will follow-up with a proposal to decrease LTs and LTVs on the stablecoin E-Modes on the different deployments.
Following the community discussion thus far, Gauntlet provides the below options to the community. While we generally aim to provide a shorter list of options, in this case, given the diverse opinions of the community, we provide a greater set of options to encapsulate the comments above. We are appreciative of all the community involvement here.
In this case, it can be difficult for the community to make a decision without concrete figures, context, and expected impact (e.g., liquidations to users) of the different options. We hope the data and explanations below help clarify the tradeoffs for the community to make a thorough, transparent decision.
We aim to move expeditiously while allowing enough time for community discussion. Therefore, we target the below timeline:
We will initiate Snapshot vote on Friday, 3/24/2023, with ample time for voting.
Following the community’s preference via Snapshot, publish AIP on 4/10/2023 to give users enough time to adjust positions as needed.
As a reminder, these parameter recommendations are for V3 Avalanche, Optimism, Polygon, and Arbitrum.
Option 1: (Include all current assets) - 87% LT, 84% LTV, 2% LB.
We recommend AGEUR to be removed from Polygon E-mode due to the AGEUR depeg and inability to repeg. However, for simplicity, we recommend an initial step of changing the LT and LTV parameters and then following up later to remove AGEUR.
Option 4: Completely remove stablecoin E-mode from Aave V3
As @bgdlabs mentioned above, certain actions, such as removing assets from E-mode, can result in liquidations. For transparency, we have compiled approximated forced liquidation data here to summarize the impact of these options. To briefly summarize these,
Total Collateral Liquidated
User loss to Liquidation Bonus
For illustration, we summarize the results for Optimism on Option 4 below. The other chains and options followed the same methodology.
Optimism on Option 4
~130 liquidatable accounts
~$2.0M in collateral liquidatable [$2.1M in collateral supplied]
~$98K collateral lost to liquidation bonus
Here is the breakdown of the top ten accounts contributing to losses from liquidation bonus:
The top user 0x8c0fcf914e90ff5d7f2d02c1576bf4245fad2b7f is a recursive USDC borrower and contributes $32.7K in bonus paid to the liquidator if the position is liquidated.
What is E-mode meant for?
Stablecoin E-mode presents a tradeoff between leveraged capital efficiency and enhanced fee generation versus overexposure to tail depeg events (as we saw with the SVB incident). The magnitude of the next black swan depeg is impossible to predict. Likewise, the insolvencies that occur at that future time will be dependent on the future loanbook and future liquidity are also impossible to predict. Indexing on the handful of past depeg events does not really capture future potential events and how dangerous they could be. The day-to-day variance and noise as stablecoins move around $1 (or other benchmarks) is natural and does not contribute to insolvency risks.
Ultimately, it is up to the community to determine how emode should be used.
Option 1: If the community wishes to enable various trading strategies, as noted by @sakulstra and @ghostlyenergy, it should pick option 1, as option 1 includes more assets.
Option 2: If the community wishes to facilitate methods to borrow historically correlated assets, it should pick option 2, as option 2 assets are more heavily correlated.
Option 3: Likewise, if the community wishes to provide a higher margin of safety in addition to option 2, it should pick option 3 instead. For example, if the community wishes to discount the future correlation of DAI with USDC per the predictions above noted by @WintermuteGovernance and @fig, the community may wish to choose this option.
Option 4: Finally, if the community believes the tail risks of stablecoin emode outweigh having the above options, it should pick option 4.
Initiate Snapshot vote on 3/24/2023.
Following the community’s preference via Snapshot, publish AIP on 4/10/2023.
DAI and USDC returns are highly correlated. If the community wishes to heavily index on the past black swan event, then the community may wish to set an even lower LT and LTV to accommodate buffer between the next black swan depeg event. The next depeg event may be even larger in magnitude.
The reason why it is difficult to simply toggle emode on/off is the forced liquidations that will occur. See the above spreadsheet for detailed liquidations breakdown at each option.
The spread between LTV and LT is less effective at mitigating insolvencies than LT. As you’ve noted, it is easily bypassed and is primarily meant to help less informed users mitigate liquidation risks and is a UI consideration. The spread is a function of pairwise metrics including correlation.
Thank you for bringing this point up. In the end, Gauntlet can only perform analysis on past data, and it is impossible to predict exactly how the composition of collateral that will be allowed to back DAI will evolve after the SVB incident, which will necessarily affect the relationship between pairwise returns in the future.
Chaos Labs is supportive of option #1, which aims for a target LT of 87%, as we stated in our Methodology post. Additionally, to further protect retail users from liquidations, we propose setting a slightly more conservative LTV at 80%, keeping a wider margin from the LT.
Finally, if the suggested option is approved, we propose removing non-USD pegged assets, specifically agEUR and EURS, as they are unlikely to be mean-reverting or correlated. We agree that for simplicity, this decision could be made following the proposed changes above.
If the community does choose this option, it is important to provide alternative approaches for achieving the desired LT while having minimal impact on existing positions.
To address this, we suggest three alternative paths, each accompanied by data for each chain:
A 10.5% reduction (below is the data on liquidations for this reduction)
Based on the above information, we recommend implementing three consecutive 3.5% reductions to mitigate the impact on liquidations and user experience. Note: Option 2 (A 5.5% reduction, followed by another 5% decrease) could be viable for Optimism and Arbitrum, but we prefer to be consistent with parameter settings across all chains.
Adding to our previous post in this thread, @sakulstra and @ghostlyenergy make good points about what e-mode is actually used for.
We are in favour of Option 1 and agree that for users to use e-mode, it just needs to offer more capital-efficient parameters than regular borrowing. This is achieved under Option 1 while still offering various trading strategies.
Our recommendation for a reduced LTV was not included in the Snapshot. According to the collaboration framework agreed upon, both risk providers carry out independent analyses. If a consensus is not reached, the community has the choice of selecting between the proposed recommendations.
Therefore, we ask Gauntlet to modify the Snapshot to include our recommendation for a lower LTV of 80% in option #1. This is to further protect retail users from liquidations by keeping a wider margin from the LT.
Furthermore, as we stated in the previous comment, the community should determine the steps to reach the LT target if this option is accepted.
Given the nuances, Gauntlet would like to clarify that we do not recommend Option 1A. Option 1A is not only less attractive to users given its low capital efficiency (it is even less capital efficient than stablecoin parameters for regular borrowing), but it is also unlikely to achieve lower protocol risk than Option 1.
Option 1A’s LTV is even less capital efficient than parameters for regular borrowing. For efficiency mode, the parameters should be more attractive than regular borrowing if the community wants to keep e-mode.
LT (not LTV) is the parameter most relevant for protocol risk. Having the same LT but a much lower LTV reduces attractiveness to users and is unlikely to decrease the protocol risk of bad debt. As such, we do not recommend these tradeoffs as optimal for the protocol.
Unless data to the contrary is provided, lower LTV is unlikely to “protect” users from liquidation.
The proposed LT-LTV spread of 7% in 1A would be higher than the USDC, DAI, and USDT LT-LTV spreads on all sidechains (which is ~5%). 1A’s spread is also higher than the spreads of all blue chip assets (WBTC, WETH). The parameters initialized for USDC on Ethereum v3 also have a very low LT-LTV spread (currently 2%). As such, Option 1A can meaningfully hurt the attractiveness of E-mode.
Here are several more specific instances where Option 1A is less capital efficient for the LTV than even regular borrowing. Option 1A’s LTV is 80%.
USDC LTV is 82.5%
USDC LTV is 82.5%
USDC LTV is 81%
Gauntlet reiterates that LTV settings can easily be sidestepped to put on a riskier position up to the LT. In one sense, extreme LT-LTV spreads can hurt user experience, should users be looking to take on debt at the LT, since they would have to create an extra transaction to withdraw excess collateral.
Gauntlet’s recommendations of the LTV for E-mode are such that the LT-LTV spread is at least slightly more capital efficient for E-mode than regular borrowing, which is why we recommended 3%.
Toggling LTVs lower has minimal impact on the risk profile for the asset. We provide the below analysis and data to show this.
For instance, Aave v2 USDC changed the LTV from 87% → 80% and LT from 89% → 87% on 2022-12-26. The below charts show the USDC balances for users grouped by user LTVs (rounded up) on Aave v2 on 2022-12-10 before this change in LTV, and the USDC balances for users on 2023-03-01.
We can see that a large majority of USDC supply does not borrow remotely close to the LT or the LTV. But if you observe the aggregate behavior of users supplying USDC most likely to be liquidated (with LTV between 80-85%), their positions have not changed at all, suggesting that the LTV decrease did not convince the riskiest users to unwind their position. In fact, USDC balances from moderate LTV borrowers (between 70-75%, or users less at risk for liquidation) flowed towards lower user LTVs, suggesting that these users unwound their borrows, hurting overall protocol capital efficiency. As such, Option 1A’s lower LTV is unlikely to decrease the risk of bad debt, and also unlikely to “protect” the relevant users from liquidation.
As a side note, reducing LTs from 97.5% to 87% requires careful implementation to mitigate the impact on liquidations and user experience. Gauntlet is happy to provide a plan moving forward on how to best structure smaller LT decreases to achieve the full reduction if Option 1 is chosen by the community.
TL;DR: Being conservative is important, but for the reasons above, Option 1A reduces E-mode attractiveness without further mitigating risk. User experience is important - Option 1A’s LTV reduction reduces the usability of E-mode. As such, we view Option 1A as the least attractive option out of the options presented.
The discussion around the proper LTV is secondary to the community decision between (1) a more inclusive E-Mode with a lower LT and (2) a less inclusive E-Mode with a higher LT.
Since the Snapshot ends tomorrow, we propose that the votes for options 1 and 1A be combined and treated as a single option, reflecting the community’s preference for a more inclusive E-Mode setup. The determination of the appropriate LTV can be addressed at a subsequent stage.
Our primary consideration remains to enable the inclusion of all USD stablecoins while minimizing the likelihood of liquidations for retail users through the margin between LT and LTV while safeguarding the protocol with more conservative LT.
Note: We stress that we don’t believe EURS or agEUR should be in E-mode for any market.
Regarding the comments above:
We have observed that Gauntlet is utilizing selective data and cherry-picking examples to discredit Chaos’ analyses, which can be misleading. In the interest of transparency and impartiality, we aim to address each point in a balanced manner to present an objective proposal to the community, thereby enabling voters to make informed decisions. It is essential to remember that this vote should not be perceived as a competition between the proposals by Chaos and Gauntlet but instead as a choice between alternate methodologies. Our objective is to present our proposal equitably and unbiasedly, allowing the community to make informed decisions.
We would like to address several points from your previous comment:
We must dispel some misleading statements regarding regular borrowing on Aave v3 sidechains. In E-mode, LT and LTV are shared across all assets. From a risk management standpoint, we must consider the LTVs of the riskier assets for option 1A, which includes all stablecoins. Please see the examples below:
Making capital efficiency or risk statements based upon the risk parameters of the safest asset, in this case, misses the full picture and functionality of E-Mode. Option 1A still significantly improves capital efficiency for riskier assets without sacrificing appropriate risk management practices.
We are in agreement that conservative LTs are crucial to risk management on Aave. However, the importance of setting sensible LTVs cannot be understated. The below graphs are from Gauntlet’s report on Aave. It is worth noting that in the report, Gauntlet acknowledges the significant impact of LTV on the net insolvent value of an asset, as demonstrated by the graph below:
The spread is crucial as it creates a buffer in which retail users are not liquidated, especially those not actively managingtheir accounts. This is also acknowledged in Gauntlet’s report and is a relatively straightforward concept.
Note: We acknowledge this report is a few years old, and we do not leverage its results or methodology in our analysis. However, it is relatively straightforward that the LT-LTV gap provides a buffer for users, both by preventing liquidations and, as Gauntlet states, by providing users with time to update their positions:
For instance, if we have a stablecoin-stablecoin loan (e.g., USDT as collateral to borrow USDC), then providing a buffer (liquidation threshold > LTV) gives a borrower more time to recollateralize a loan. (https://gauntlet.network/reports/aave).
Apart from providing users with a time buffer, we also protect them against depegs of a size approximating: 1 - LT/LTV. Our spread of 7% protects users (borrowing at or near the LTV) from an up to 8.75% deviation in stablecoin prices, whereas a spread of 3% only protects up to a 3.57% deviation. The math underpinning these percentages can be found in our paper, as well as graphs on when deviations have historically exceeded 3.57%. In our paper, we only considered DAI, USDC, LUSD, and USDT (current Ethereum tokens)
Gauntlet has also alluded to the importance of LTV-LT spreads in the Q&A section, in their response above:
This comment misses the point of lowering LTV. We suggest using a lower LTV to set guard rails for less sophisticated users, which comes from a concern for the users of the Aave protocol, not as a means toward avoiding Aave protocol bad debt. Furthermore, the fact that users borrow at lower LTVs than the maximum LTV for USDC is not an argument in favor of increasing the LTV, similar to how we do not increase the speed limit on the highway simply because drivers are driving well below the current limit. Much like LTV and LT, the speed limit is meant to prevent users from taking more risk than is sustainable according to relevant risk metrics. To set this parameter, we consider the case of if/when users borrow at or near the LTV.
Considering the volatile nature of the DeFi space, an LTV of 80% and an LT of 87% represent a substantial enhancement in capital efficiency for most stablecoins, particularly those currently isolated and possessing low LT/LTV. This approach allows a more inclusive E-Mode, mitigates bad debt via a low LT, and safeguards retail users by incorporating a sufficiently wide spread between LT and LTV.
Let’s take a step back and find common ground. I want to acknowledge your analysis and your team’s effort. Our team offered a different perspective based on alternative methodology (as you mentioned above). As we continued the discussion and were tracking the active vote, we realized that there might be additional information that could benefit everyone in the community. Let’s work together to explore those possibilities.
The ACI previously voted for Option 3 as it was our considered most balanced option between being more conservative and having less overall impact on existing users.
Due to recent debates, we switched our vote to NAY, not because we don’t think there net value in the proposal but because recent discussions pointed out that there’s might be a benefit to taking a step back and continuing discussion.
We will be happy to cast a vote in another snapshot, when the proposal has achieved more maturity.
In light of the community discussion Gauntlet plans on putting up another Snapshot vote in order to find common ground.
This new Snapshot vote will have the same options except that Option 1 and Option 1A are consolidated into a single option where the LTV is 82% (in between 80% and 84%).
While Gauntlet recommended 84% for Option 1, in order to reduce governance fatigue without sacrificing the core principle of the option, we present this updated proposal as a common ground, interim solution.
If this option is chosen, the community can monitor how usage evolves, and can then adjust the LTV accordingly. For example, depending on how attractive this new E-mode is to users, the community may wish to adjust the LTV upwards or downwards. Fortunately, adjusting LTV has significantly fewer downsides than adjusting LT, as LT can cause liquidations while LTV cannot. As such, the community can be more dynamic around adjusting LTV. Again, we note that the main parameter behind protocol risk of bad debt is LT, not LTV.
We plan on putting up this updated Snapshot next week, with ample time for voting. If @ChaosLabs@MarcZeller or any other community members have feedback on this consolidated option, we’d appreciate hearing your thoughts. We thank the community for their contributions on this proposal.
As previously stated, we view the discussion around the LT-LTV spread as secondary compared to the primary decision of what type of E-Mode should be implemented - (1) a more inclusive E-Mode with a lower LT and (2) a less inclusive E-Mode with a higher LT. Therefore, we support this interim solution to move forward quickly with an updated Snapshot.
To improve our governance process and prevent similar situations in the future, we would like to emphasize the following best practices:
Post the original Snapshot only after discussions have matured and all feedback and analyses have been conducted.
Collaborate to deliver a clear, unified message to the community, incorporating all relevant parties’ arguments, in case new material information arises during the Snapshot.
Snapshot vote has been published here. We thank the community for their participation.
We note that Option 2 and 3 are identical, except that Option 3 has more conservative parameters.
As mentioned above, for Option 1, we are considering removing AGEUR from Polygon E-mode due to the AGEUR depeg and inability to repeg. However, for simplicity, we recommend an initial step of changing the LT and LTV parameters and then following up later to remove AGEUR.
The Snapshot vote has closed, with Option 3 being the community’s preference at 532k votes. We thank the community for their participation on this impactful vote for the protocol.
We aim to move forward transparently and give users enough time to adjust their positions if needed. We provided liquidations data above, and will provide regular updates on liquidatable accounts during the AIP phase.
To give users ample time to adjust positions, we target an AIP on 4/24/2023.