Hey, sorry I missed this topic earlier. Yeah, jump risks and slippage are the two main source of losses. Just to reiterate what I was saying on the call, imagine if the asset price moved continuously (no jumps), and at the liquidation price you could auction off infinite collateral. It would be impossible (I hope) to construct a loss in such a scenario.
In reality, we know (and observe) that asset prices do make sudden moves, and that auctions can lead to slippage. The high level approach to the risk model is a Monte Carlo Value at Risk. Using a MC simulation is nice because you’d like to incorporate some path dependency (concepts such as user behavior, i.e., dynamic exposure amounts), but more importantly, historical sampling is just terrible in the space. Every asset looks amazing at 150% LR, even the shitcoins. Using MC allows you to be a bit more flexible. (I’m referring to historical sampling of the asset price, but same with the auction history, which is extremely short and unusable.)
Honestly, modeling both jump risk and slippage are extremely tricky. In many ways, our approach should be viewed as an MVP, a tool to get us through the early days. The “right” approach would require significantly more resources than we have at our disposal (just consider the fact that major institutions of thousands of risk analysts). But I digress. For jump risks, we’ve taken a look at as much history as possible (looking at ETH prices from pre-SCD) as well as looking at Bitcoin. We’ve also done analysis on looking at some shitcoins that have succumbed to event risk.
Our insight with jump risk is that, at least some (although I will argue a majority) of jumps are event-driven and can be profiled. So we began to categorize the causes (exchange hacks, regulation, major hard forks, protocol hacks) and tried to characterize a) how frequently they occur, and b) how severe they are. So we can look at an asset and say “hey, this thing is really likely to get hacked because we know its been engineered terribly” and from other similar hacks we know that the token is likely to drop 50% in such a scenario. We tie the fundamental analysis of an asset directly to its likelihood and severity of its jump. Now, this is of course highly subjective, which is where we’ll be leaning on community input (ultimately they have to be OK with the risk parameters). But our model can help reflect the community’s viewpoints.
At the technical level, I agree there isn’t a lot of good literature on stochastic modeling of jump risks (you really have to be a super-quant to have a nuanced understanding of the differences between various approaches). I would argue that at this early stage, with the limited amount of data we have, it doesn’t really make a difference. Much of this will be practical sense. We spent some time exploring other models, and would love to have some other knowledgable people participate here. Check out this paper, for example. This is probably the best paper that most closely represents our approach https://arxiv.org/pdf/1604.05404.pdf He uses a pretty sophisticated stochastic model with jumps, one that is probably more than we need.
Hope this gives some good context for now. Would love some comments / feedback. I’ll try to share some data around the jumps tomorrow (cc @Primoz ). Will also try to do another post on slippage. Lastly, please don’t expect us to do anyone’s homework for them haha! We’re hoping this will be a community effort, there is ample room for improvements and iterations.
Thanks for all the discussions you’ve sparked!