VaR Deepdive
tl;dr We have updated our Risk Dashboards and have split our existing VaR into insolvencies (new definition of VaR) and liquidations (Liquidations at Risk or LaR). As a result of this change, VaR will now appear substantially lower and the bulk of capital at risk will appear in LaR.
Over the past couple of months we have been gathering feedback via user studies, surveys, and emails about what parts of the Gauntlet Risk Dashboards are useful and what parts can be improved. We have gotten a ton of great feedback (thanks to everyone who gave it) and have been making changes to improve the usability of the dashboards.
The largest feedback we received was around our VaR statistic: what it was and how it should be interpreted. We identified two key problems with the statistic:
- It combines liquidations and insolvencies together into one metric when it should be two
- It isn’t clear how to interpret the metric in the context of our recommendations
The impetus behind our current formulation was to present a simple understanding with only two system metrics: VaR for Risk and Borrow Usage for Capital Efficiency. We thought this would help users rationalize the tradeoff we are making in a simpler way (i.e. VaR and Borrow Usage tend to increase when increasing capital efficiency and the opposite when reducing risk). However, a major part of the metric — the fact that there was very low insolvency risk — wasn’t clear; the metric was dominated by liquidations that would be safely absorbed into the market.
We are changing this and are breaking out liquidations from VaR, which will be presented as a new disjoint System Metric in our dashboard called Liquidations at Risk (LaR). This will be a big departure from the current definition of VaR, as the dashboards will now show much lower numbers in terms of pure insolvencies. It is important to note however, that this is purely a UI change. Our parameter recommendation methodology (which we will follow up with in another blog post), already makes a distinction between Insolvencies and Liquidations. The Insolvency risk has been on the dashboard on a per collateral level since the initial launch; it is represented both in our Net Insolvent Amount heatmap as well as in the Sim Statistics section.
Why VaR?
At Gauntlet we are primarily concerned with the risk of adverse tail market events (i.e. major price drops in ETH and other crypto assets). While these events are infrequent, they end up being the highest risk market events for lending protocols and collateralized DeFi in general.
The biggest risk to a lending protocol is insolvencies; this occurs when a protocol’s debt (borrowed assets) exceeds its assets (supplied collateral). When this happens protocols need to pay off the underwater loans, which each protocol achieves using different mechanisms (i.e. Aave’s Safety Module). Note there may be situations where a given protocol is comfortable with some level of insolvencies (e.g. $10,000 of insolvent debt can be trivially covered by $100M in the treasury or insurance fund).
Value at Risk is a well defined metric commonly used in banking and investment portfolios. It quantifies the capital that a protocol (or company) may lose during adverse market events (or that is at risk in general). It is a measure of tail risk and can be calculated in multiple ways. For our metric, we leverage the Gauntlet Platform and specifically focus on running Monte Carlo simulations to approximate this tail risk.
It is important to note that VaR is sensitive to model inputs; it will change day to day due to changes in volatility, user borrow positions, and asset correlation structure. In addition, VaR tries to estimate long tail and infrequent events, which by definition, have higher variance as a prediction problem. As such VaR serves to help one broadly understand the risk in the protocol, but does not describe the market risk profile in its entirety. To get a deeper insight into the risk profile, you can look at the market and sim statistics on a per collateral level in our dashboard. This will allow you to inspect the entire distribution of insolvent debt and liquidated collateral (via our heatmaps), as well as specific market inputs and simulation outputs.
What is changing?
The primary difference in our dashboards is that the VaR statistic is now the 95th percentile net insolvency amount that we observe at the end of our simulation runs when modeling across a range of volatilities. We are also adding a new metric Liquidations at Risk (LaR), which is similar to VaR in construction but captures potential tail liquidations instead of insolvencies.
Let’s break this down in more detail.
Net insolvent amount
This refers to the amount of capital that is found to be insolvent (unable to be liquidated for a position that is below the required collateralization ratio) at the end of a simulation run. Technically, an account i is insolvent if the total value of debt D_i is greater than the total value of collateral C_i owned by the account. The net insolvent amount is defined as the total value of debt minus the total value of collateral, summing over all the insolvent accounts.
For a collateral asset k’s net insolvent amount, we calculate the sum of the pro rata insolvent amount from all the user accounts.
where c’_{i, k} is the user i’s pre-sim collateral amount for asset k and
is the user i’s pre-sim collateral across all assets.
Note that an account’s insolvent amount is calculated at the end of sim and the pro rata distribution of insolvencies is based on pre-sim values.
95th percentile
We run thousands of simulations across varying levels of volatility and the risk parameters in the protocol in order to assess how various rational agents will react under different conditions. After running these simulations for a given set of parameters (i.e. the current parameters in the system, or the parameters Gauntlet is recommending) we then order them by the net insolvent amounts and then take the 95th percentile value.
Range of volatilities
Our risk analysis serves to better understand what happens during tail market events, as this is where protocols are at the biggest risk of insolvency. This means running our simulations with greatly increased volatility to simulate these tail events (e.g. major price drop). In order to forcefully increase volatility we multiply the asset’s current (28 day annualized) volatility by a Volatility Scalar.
For example, a price drop of 50% in a day would roughly correlate with a 1400% volatility scalar if the asset’s current volatility is 100% (right now ETH’s 28 day annualized volatility is 79%, COMP’s is 128%, and AAVE’s is 141%).
As you can see in our dashboard heatmaps, we run simulations over a range of volatility scalars from 100% to 1500%.
What about Liquidations at Risk (LaR)?
Liquidations at Risk are very similar to VaR in formulation, except instead of measuring the net insolvent amount, we instead measure the amount of liquidations at the 95th percentile.
Liquidations, while healthy for the protocol, can adversely affect Borrower UX. Specifically, when lowering Collateral Factors/LTVs/Liquidation Ratios for given collaterals some users’ could be forcefully liquidated. Right now the systems in place to inform users of lending protocols to ‘top up’ their collateral are minimal to none. The first time we made a change to reduce CFs, we had no way of informing the users who were about to be liquidated as a result. We tried to send ETH transactions with a message telling them they were going to be liquidated, but observed no response or action on behalf of the users.
It is important to underscore that liquidations (that don’t cascade) are healthy from a protocol perspective. They are what allows a lending protocol to stay solvent. This means in times of increased risk, there may be scenarios where we must reduce CFs/LTVs (and thus potentially force liquidate users) to ensure the protocol stays secure. That said, we understand that the lack of tooling right now makes this a somewhat suboptimal experience, and as such we don’t like to churn CF/LTV parameters too frequently. We will expand more on our parameter recommendation methodology in a subsequent post.
What about cascading liquidations?
Cascades of liquidations (as discussed in detail in our older Aave and Compound reports) can impact external market prices which in turn lead to further liquidations (i.e. a deflationary spiral). This can both exacerbate the effect on asset price (sending it downwards) as well as making it potentially unprofitable for liquidators to liquidate more positions. Liquidations like this are captured in LaR up until the point they can be absorbed by the market and then VaR for the portion that is insolvent.
Summary
We hope this post helps the community better understand our system metrics and the reasoning behind this change. We plan on building more features into the dashboard as we improve our models and refine our recommendation methodology. We are always keen on receiving feedback from the community. If you have any feedback please feel free to reach out via the Send Feedback button on our dashboard, or directly to dashboard-feedback@gauntlet.network.
Special thanks to Hsien-Tang Kao and Nathan Lord for their help on this post
