Galaxy digital - Underwriting in DeFi Study of Current Practices

Transcript

all right uh everybody we're gonna get started thanks for being here and i hope you enjoyed the lunch break so we'll we'll talk about underwriting and defy and will is from galaxy digital stage is yours thanks hey everybody um so quick thing there's a little bit of a conversion challenge from powerpoint to google docs so that's why this so i think two two slides will look like this um this is not probably the talks that other people have given i apologize for that in advance i took the opportunity with this presentation to study a topic and look at a topic that i've been meaning to look at for quite a while

underwriting in d5 so this is going to be a little bit more focused on a particular problem than on the general vc landscape but i realized this you know way too late to change it yesterday so what i'll do is i'll try and run through this really quickly and then we'll open up time for you know questions at the end where i can talk more discreetly about you know sort of financing deal structure and crypto and token rights and equity versus network and other things like that um i'll also note that none of this is financial advice obviously and um and none of it represents the sort of opinions of galaxy digital broadly um mostly this is me and i and it doesn't

represent the views of the of the company in general so what is galaxy some of you may know it some of you may not galaxy digital is a large financial um um you know firm in the crypto space um with a balance sheet of a couple billion dollars of assets and hundreds of of institutional counterparties so hedge funds family offices um non-profits often trade through us primary businesses are our otc trading desk um plus plus um quant quant trading lending banking asset management um mining i sit in the principal investments group at galaxy which in practice is really just the venture capital arm of galaxy digital and our our mandate is to go find the problems

in the space that are to you know and solutions in the space that are two to five years away from their maturity invest behind them build relationships with the founders so that once those things come to maturity the rest of our businesses orient around around these solutions we have a big operating business in the space of 350 people and so one of our real advantages is we often identify problems before they're you know before others in the space too so identifying in 2018 that sending you know settling otc trades on ledger wallets was a terrible way of doing doing things allowed us to figure that fireblocks was actually quite a good way of doing things so we became an investor we became one of their first customers and

you know they scaled from there so as i said i'm really looking to talk about something in particular that i've been you know thinking about and leaning into recently d5 remains a core part of galaxy digital's interest and my own interests i think this is a particularly good time to re-evaluate the purpose of defy as well as reevaluate the advantages and disadvantages for one it proved really you know it proved really robust during the the recent volatility versus traditional lending funds um well it has some limitations you need to onboard better assets the the model of you know a random on-chain asset that represents trading on an exchange or equity over an

exchange that where where assets are transacted price rising yields rising you know usage rising price right it's a it was a circ it was a nice feedback loop in 2020 and 2021 and obviously we're seeing the the downside of falling out of that feedback loop in particular i want to focus on underwriting very few people talk about the concept of underwriting and defy so i'm going to talk about what is underwriting um underwriting in collateralized lending and what what what's actually going on there um underwriting and other types of lending in d5 and then the future you know potential future directions for underwriting including things like maximum likelihood estimation so underwriting is essentially just the

pricing of risk right and what all all it means and there's three types of underwriting typically there's loan underwriting there's insurance underwriting and there's investment underwriting today i'm going to talk primarily about about loan underwriting because it's the most relevant to d5 loan underwriting really involves the modulation of the modulating of collateral interest rates repayment repayment schedules um loan covenants and so on and so forth to adequately accept risk and onto a balance sheet what that in practice means is taking a multi-variable multivariate design space or a multivariate parameter space including borrowers financial history their financial position uh your you know as a you know your particular loan

portfolio and collapsing it into primary into really two things ideally um probability of default and loan and the loss given default those ultimately you know together you know provide an expected loss on a on a loan position um so how where do we see this in in d5 today like we nobody ever talks about what it means to underwrite risk in d5 and i think that's part of the problem um we see it a ton in over collateralized lending we never we don't talk about it as underwriting but when you define a liquidation threshold or a risk premium or an interest rate you're defining a you know you're doing underwriting so governance and maker is doing underwriting uh and one of the things you notice is there's three eth vaults

right there's an eth vault with 145 collateral an ethol with 130 collateral vault with 170 collateral each of those has different interest rates there's underwriting going into that that determines what the fair interest rate for a particular you know type for a particular collateral level is so every maker loan is embeds a probability of default and a loss given to fall so if you take a look at a 180 a vault that has current 180 collateral ratio with a 40 145 liquidation threshold and you sample the returns the hourly returns from the last year of uh ethereum ethereum trades or ethereum price action

you can model a monte carlo simulation that gives you 62 probability of default on that particular vault now the probability is in practice much lower because the probability that somebody tops up their vault is is actually you know pretty good and so in practice fewer than 62 percent default but if you just relied on the volatility hitting the 145 threshold um you would end up with a lot of defaults but the important part of of of underwriting you know over collateralized loans in d5 is that there's a zero percent loss given default right that's at least that's the target the idea is if somebody defaults the collat the liquidation mechanism will actually handle that um

handle that debt and resolve that debt before it filters down into a loss for the maker protocol so it's interesting how even little things factor into the underwriting process the liquidation the switch from liquidation 1.2 in maker to liquidation 2.0 in maker actually just reduces the loss given to fall by reducing the amount of capital that's required to put up in a collateral auction and going from an english auction to a dutch auction reduces the loss given to fall so that's a that's a basic form of underwriting that maker governance is doing but what about other forms of lending so one of the only other forms of lending that's actually going on in practice on chain right now is real world lending um

there's a couple others but i'll focus on real world lending because i think it's the most present so collateralize lending real world lending so these are typically these are uh underwriting is more complex you're looking at loss given default of non nonzero loss given default and non-zero probability of default so there's four protocols that are primarily doing this um forgive me if i'm forgetting any because there are a couple couple more um maker maple centrifuge and and um goldfinch they all at least have two plus assets on the balance sheet that are real world real world loans it's a growing space um we've you know even in even in a time of sort of market distress we're growing up to you know the total

total um amount in uh originations and in sort of i sorry incapacity for real world assets is about 900 million right now almost a billion so um it's it's a sizeable problem to to to talk about and so it's worth analyzing how they're actually doing the underwriting and how risk is being sort of accounted for and and then priced on in these protocols um one case study is the maker huntington valley um loan so i forget forgive you guys if if you didn't read the whole proposal because it was quite long i did and i'm gonna i'll bring some notes for you so huntington valley bank is a commercial bank in in pennsylvania and they entered into an agreement with

maker where maker forms a trust called the maker dow bank participation trust i'm just going to call it the maker trust which is a delaware trust that benefits maker dow the delaware corporate trustee verifies that all actions on behalf of the trust are carried out in accordance with maker votes and maker issued a hundred million dollar debt ceiling to this maker trust so the actual structure of the of the agreement is that maker is maker participates in all loans uh with uh hvb pari pursue which means they basically sit as a side car and get the same terms um h the require the requirement of the deal is that hvb retains greater than 50

ownership in all loans uh that they do so um it ensures that they have sort of adequate you know sort of risk in the in the loans that they're essentially putting on maker dow's balance sheet and pricing on behalf of makerdale um so it's clear who's actually underwriting this risk it's not maker it's hv bank and there's a decent amount of structuring involved to make sure hv bank is maximally aligned with the interest of the maker token holders why why does maker want to do this well it gets them access to really good underwriting for free as long as they sit as a side car to hv bank why does atv bank do it well hb bank actually has a seven million dollar max

capacity that's legally mandated on per customer basis so they can only give out seven million dollars in loans to their customer and so you easily see that this allows them to scale up what they actually can extend and credit to their to their customers um another protocol that's doing some underwriting is maple um maple is a protocol that facilitates a particular pool delegate to to spin up a balance sheet using depositors money essentially investors who provide capital to the balance sheet of a pool and then the pool delegate actually um sources underwrites and then executes on a particular loan that then the the

investors benefit from the interest rate and the pool delegate takes a fee pool delegate you know partially partially capitalizes in a junior in uh like a junior pool or a first loss pool which is you'll see this in centrifuge as well um but here it's clear too that and uh the person who the underwriting is happening as a result of a century of a third party here the maker or the maple pool delegate so i'm not making any sort of sort of arguments about that i'm just stating it as it is and maker is makers providing capital or in the process of providing capital to maple through their d3m which is their uh dye direct deposit module i think is the name of it and then there's

centrifuge which ultimately brought me to this problem centrifuge is a is a protocol that allows asset originators um to go interact with real world borrowers enter into loan agreements with real world borrowers and take that debt that they buy and then stuff it into an spv that spv is then securitized by centrifuge so what that means is centrifuge produces an additional set of assets that basically carry the cash flows of the spv along with them and that's all happening on chain what's interesting about this is that is that makers buying the collateral so the senior tranche of collateral the the one that's protected by the law the losses of the junior

tranche is actually used as collateral in maker and then makers recycling die back to fund the underlying borrower so if you see the the mouse here the bar is going to the asset originator the asset originator is agreeing to debt and then originating it putting the debt which is an asset into an spv then the funds the sort of prof proceeds from the spv are getting paid out to maker and then the and then die is getting returned to the borrower now what's really interesting about this is that the waymaker actually underwrites these deals is that all they do is they just evaluate the asset originator they don't actually check the loans that are in that are that are in the spvs so they'll give the asset

originator a debt ceiling say 20 million dollars and then the asset originator can put whatever debt they want in that pool which you can see where that gets challenging right now it works okay because again there's structuring going on that make sure the acid originators are maximum maximally aligned with with maker's interest by you know forcing them to to put to hold junior trunch assets so if there's losses they face some losses too but it doesn't really protect the problem if you know an asset originator just wants to warehouse the best debt for themselves they they can do that without putting you know without putting that really high quality debt in in you know the spv for me that's used as collateral and maker so you see how this

kind of all becomes an issue um what you realize when you actually go down this road is that protocols are are autonomous balance sheets functionally i think this is especially true in a world where compound and ave are have released sort of designs for stable coins and are going to pursue stable coins or some sort of debt-based asset um assets come in and they you know form there those are the assets that the that maker something would purchase that would yield to maker and then the liabilities are die or or the compound stable coin or the obvi stable coin in that case but it's very clear that even though these are autonomous balance sheets they're extremely limited in their capabilities for pricing risk and that's a huge problem when it comes

to actually you know financing real things that are not just zero percent loss given to fall so um we have essentially two cases that we've discovered um easy risk i.e where the target loss given to fault is zero percent can be done programmatically sort of arbitrarily by by the sort of smart contract hard risk right now is being outsourced to a central party because maker or compound or or xyz or maple have no way to actually price that risk themselves without just relying on somebody a central party that's sort of maximally aligned with with the underlying balance sheet um to price it and if you think this is is a

specific to real world assets um it's not i encourage you to go look at um the um porter finance mip6 proposal port of finance does dow convertible bonds and they're all also doing the underwriting themselves so i mean it kind of poses a question like is defy good for um is all if it can't price complex risk is all it good for is just capital formation and then you have a centralized provider that actually and analyzes the risk and all we did is just put funds in a pool for that centralized provider to to sort of allocate may i hope the answers i hope the answer to that is no and i'd like to get to a place where we can price more complex forms of risk so

i started looking into this concept of decentralized underwriting how do you actually you know underwrite using um not a central party but a sort of generalized primitive or infrastructure can we bring a new primitive into into being this is a lot of what i work on as an investor we want to you know invest in and support new primitives whether that's companies or protocols and so you know this is one instance and i hope you guys realize from this there's probably more primitives that are still waiting to be discovered and um so that's you know that's what gives me gets me excited um i spent some time working on with centrifuge on this problem um the fundamental question we started with is what would rating what

would a rating on chain rating agency looks like look like so if there was a sort of gatekeeper that sat in between the asset originator and the maker and the spv that's maker collateral that's determining what loans get to go in the pool or not in the pool right um so one way you could do this is really basic like stake weighted voting on particular loans and if you sort of uh create a profit motive to stake weight behind good loans i.e you profit if the loan is ends up being good and then you you you you sort of lose money if the loan ends up sort of going underwater and you get use your stake at slash like you can imagine a um a sort of decentralized underwriting system or a decentralized ratings agency

emerging that all protocols could theoretically benefit from this is really early stage the work isn't done and we haven't come to any great conclusions quite yet um and then there's i i'd say i'll put this in people's minds directions for further research um i think there's a particularly good area of mathematics called maximum likelihood estimation for this for the study it's really the study of taking taking large large numbers of bids from from uh you know sort of like various parties and then aggregating them into the most likely probability distribution that accommodates their all their bids um i'd say using mle to to analyze probability of of default and loss given

default for any particular loan and then sort of matching it with an incentive system uh could be a really really interesting um you know future direction of research um for this sort of like on-chain ratings rating system or ratings agency um so that's pretty much it like all you know as i said what i really wanted to do in this talk was explore an intellectual direction that i was you know interested in and then um hopefully it shows you that there's still primitives to be found we've been working on and thinking about this problem for a little bit um but i know it's a little distant from what other investors may be presenting about shows you galaxy's interests my own interests um but i also know that you know there's certain questions you guys as uh

founders in this room want to get answered that were not not necessarily answered by this um presentation that are more you know related to fundraising and you know how to structure deals and how to how long they take to get done and and so on and so forth so i want to open the floor up for that um absolutely i was gonna i was gonna say that and i was gonna say i think that's also the beauty of fcc is that we have like this kind of content and i think that's super important for everyone to kind of see also what are the challenges and the research areas that are open so thanks for that and yeah i just wanted to spend maybe the last five ten minutes um and maybe you can say a word about galaxy because you're doing much more than just um venture i think that shows

through your presentation you know there's a research part there's a training side of things there is a venture arm there's galaxy digital there's the i think there's a gaming fund as well so there is a gaming fund exactly so if you can say a few words maybe about that would be interesting for the audience and then we'll um circle the mic for questions obviously yeah um as i said galaxy is a real pretty large organization we've got 350 people in the venture team there's right now there's nine but in other teams lending we do a lot of uh financing of minors so we'll give out we'll take in um you know bitcoin or eth and we'll then we'll basically give them funds to scale up their operations we're also doing some proprietary mining and staking ourselves um our our banking team i know nobody likes

bankers actually you know for founders they for our founders historically they've been very accretive um you know you know getting the ability to plug in plug in our our startups with people who can help them you know help them get aqua hired or help them get acquired or um help them you know if you reach that scale go public is you know has has been valuable to our founders and to our team in general um we do a lot of trading quant trading our quant trading team does a couple billion dollars a day on exchanges uh they're not that's not profits that's volume just to be here yeah yeah sorry of volume thing otherwise we're all gonna apply for a job at galaxy no and and as i said like in some ways the investment team has

always been the center of the the organization because there's so much that is going on in the space there's so many talented founders there's so many new ideas and new new directions that are still left to be explored and so when you have a nascent space like that that's fast moving you need a you need a team to be evaluating and working with the talent that's trying to solve the hard problem so i think what we're really good at is identifying some of what those hard problems are because we face them every day in so many different capacities and then passing them on to the founders passing those hard problems onto the founders who can who can make a difference for them and we can you know support financially and um and uh with our you know with what the company has to offer um so

that's a little bit about galaxy um i think we had a few questions yep hi thank you for the nice very interesting talk um i had a question that actually stemmed from the slide we were talking about centrifuge uh but it kind of also relates to the uh the mle the max molecular estimation um you mentioned how the major problem in centrifuge is the fact that like there's no clear incentive mechanism for incentivizing the users to to put high quality debt into in into the spv as opposed to keeping it themselves um i think there's also another factor that's super important uh which is

um the the correlation of the risks across all the debt that is that is being issued because that's we saw that we saw that obviously very recently with the the terror collapse like on like on polygon there's there's an insurance there's an insurance protocol that was uh selling insurance and it was selling insurance to a bunch of different protocols and users that were all effectively exposed to terra and so once tara collapsed every like the entire entire insurance uprise completely broke apart because you have to protect yourself from systemic risk so can you say something like what how do you what do you think is the way forward in that direction yeah i mean i that it's a great point i think i focused a lot on underwriting individual loans or particular loans but on the underwriting

process involves actually analyzing the sort of collective risk of a of an of a of a portfolio right and the the correlations between different assets in the portfolio the stan you know the standard deviations are the volatility of particular assets in the portfolio so 100 agree with the point like um you know collective risk management and and analyzing risk of an existing portfolio is as big a part of underwriting as um as looking at individual loans so i appreciate the point it's a great it's a great one yep hi uh thanks for the analysis and i think it's to me it as an investor clearly showed that at least three out

of the four models show points of weakness when it comes to stress so taking it a step forward like based on your research have you seen any any practices or any attempts where reinsurance could be added to these models as a safety net so that even semi-perfect models could actually work um down the road because if d5 were to work it needs to be trustless and it can't be in many instances so reinsurance might be a big help to do that uh what have you come across in that is yeah not much on reinsurance couple uh yeah again it's it's just it's pretty nascent uh in terms of people trying to reinsure

any of these protocol risks that you know um i talked about in the presentation today so it's unfortunate i don't have a great answer for you if somebody is working on that i'd love to you know learn more yup last question actually it's not a question it's an answer so whatcom is possiblizing precisely this product so preschool welcome we are working on this w-a-k-m-a-m it's an actual insurer uh today you have decentralized insurance meaning that it's token orders which will decide if you should be reimbursed when you are supposed to have an issue and it happened that some decided not to on some context

so you have this risk that's the purpose of let's just say a true decentralized dao at whatcom we believe it's important to have an actually an actually regulated insurance meaning that you have the french laws if you are unhappy like you you subscribe the insurer decides that you should not be paid you may sue us while you cannot sue it yeah that's uh not one thing to discuss this uh and thanks a lot for your talk yeah that's a very important subscription uh problem like who decides when you have to pay in defy in many contexts it's very unclear looking at your um monte carlo um at the

start um what sort of model did you use to progress the the processes that you looked at did you use for example a historical based model or do you use more of a kind of a market based model where you might look at a volatility surface and then derive like a a dupia local vol or something like that like super simple i i did this like sunday like late sunday night so um i just sampled from the historical distribution basically of the past year of hourly returns and then progressed the model forward based on based on an individual sample from the distribution so it's like not a high-tech monte carlo um

yeah 100 you could have in that case you could have probably priced like that probability of default not using a monte carlo simulation i think there's an analytical solution to it which is like roughly the the price of an american put option like the probability that american put option um expires in the money essentially if 145 is the strike price plus some or multiply by some probability that the underwrite or the sorry the vault owner doesn't top up their collateral um because if they top up the collateral then they don't default but if they don't top up the collateral then they are sort of liable for for default there so um i think i'm out of time great

question i would love to expand the scope of that that model but um that's part of the project so all right thank you so much thanks thanks a lot thank you amazing thanks thanks so much where should i put this uh yeah just to get her thanks all right uh so those who want to go in