Decentralized finance (DeFi) lending protocols like Ethereum Credit Guild and Ajna have opened up the creation of new, cryptocurrency based lending markets that cover a long-tail of tokenized assets. One under-discussed topic in the DeFi space is the generation of lending terms and collateral selection as an overall credit risk framework. This paper discusses a rough overview of the problem for these DeFi credit markets establishing why it’s important to new participants operating in the DeFi credit space. A second section is included that explains basic frameworks that should be used as a panel of risk assessments when considering credit markets, which should be barebones risk metrics for any new lending terms. Lastly, new avenues are discussed as possible options for credit operators that have appropriate capital, that are possibly under-explored opportunities in crypto-based credit markets. The overall goal is to explore crypto-credit markets for beginners, and generate interest in development in new products that can serve these markets.
I am also available to join risk engineering and analytics teams in developing these products for protocol operators if interested. Feel free to reach out to me.
A simplified model for lending when collateralized by volatile assets is that the system should try to optimize:
While these may seem trivial, global lending rates are a competitive market where participants may undercut other lenders to gain market share. Also, volatile collateral may rapidly move against lenders where after liquidation the debt may never be recuperated in full. The interest rate is in some sense both insurance against default (risk premium) and a payment that charges a fee for general credit (time premium). The difficulty in the cryptocurrency space is compounded due to the extreme volatility on the upside and downside of both collateral assets and lending dollars.
An additional quirk in cryptocurrencies is that despite the name, not only do cryptocurrencies as an asset class include assets that should not be considered “currencies”, many users within the space and outside of it cannot agree on whether the assets are equities or commodities.
While traditional businesses can lean on governments to enforce equity contracts and associated contractual cashflows, cryptocurrencies that use blockchain technologies generally only encode fees as income as the base “cashflow” for not only protocol applications, but blockchains themselves. From that, there are analysts who treat these system’s tokens as “equity” and model valuations based on discounted cashflow. In standard equity models, pricing cashflows under discount is seen as a sort of collateral that lenders use to assess risk of loans and default probabilities, but it is unclear whether this model works in the blockchain space.
Coupled with this, many cryptocurrencies serve a “commodity” use-case as they price abstract refinement of data, computation, or capital allocation in the future. Some analysts model systems asserting they are informational commodities who’s value is not dependent on cashflows but future refinement of goods that have downstream profits for purchasers, which adds a complex valuation dynamic in the supply chain of information processing and data storage.
Many cryptocurrencies can be argued as both an equity premium that sells a commodity, or a commodity itself that is used to refine other (financial) products. These difficulties make the models of using traditional equity risk frameworks suspect since these systems sit somewhere between cashflow generating organizations, and simple information systems.
Most protocols enable auctions for collateral liquidatation if debt is deemed bad and the system wishes to call back the loan. Problems occur in a free-market system such as Ethereum during these auctions, since there is no real obligation from participants to purchase the auctioned debt. If a lending system like a bank wishes to liquidate collateral in traditional markets, there is some contractual agreements between market-maker that they must operate and quote reasonable prices, and some legal protections in place if this obligation isn’t met. In Ethereum and decentralized finance, this obligation isn’t a requirement, and in the face of large liquidations, market makers may not re-purchase assets at a value that covers debt.
Another larger issue is meta-level effects of blockchain systems that may allow the blockchain substrate itself to be “gamed”, to allow for auction buyers to purchase significantly discounted collateral during the auction by attacking the substrate itself. While this is unlikely today, during previous large scale crashes, Ethereum and other side-chains were “spammed” by nefarious actors, preventing other auctioners to bid for liquidated collateral. Modern collateral auctions today have risk mitigation techniques to help prevent this type of attack, but this risk is present and may affect credit pricing as a risk premium lenders must cover.
A standard metric for lending rates is to take account of the daily standard deviation of percent returns, i.e. volatility. Under statistical normality assumptions, this works out to be a tractible metric to discuss and rank risk. Problems do arise as this method smooths outliers, which leads to Value At Risk (VAR) metrics as a more robust non-parametric statistic. Here, one ranks losses across a sample range and period, and can describe “top 2%” or “top 5%” losses. This method is more robust and can take into consideration of outliers without the smoothing effect typical in statistical metrics.
One consideration for cryptocurrencies unfortunately is both VAR and volatility metrics have some of the highest percent changes in assets. Under normal assumptions, yearly volatility of cryptocurrency ranges between 50 and 90%. It is difficult to generate a real discussion of lending VAR/Volatility metrics across cryptocurrencies when volatility is so high.
While these metrics in general difficult to work with in lending parametrization, they should likely be used as part of a cohort of lending factors. Also, lending operators should use these metrics as a way to identify ranks of collaterals to understand comparative risks between assets. Unfortunately in downturns of markets, many crypto based assets correllate together and down, so comparative metrics should be taken with skepticism.
Using the Merton Model derived from the Black-Scholes equations, one can produce a closed form estimate of default probability for a particular lending agreement. This model utilizes:
While the specifics number of the risk-free rate and collateral volatility can be debated, since they do not reflect respectively the collateral’s growth rate and parametric volatility is an estimate respectively, this particular model can serve as a way to guage a particular lending agreement’s margin of safety as terms are created. As well, this particular model’s tractibility allows for real time monitoring of lending parameters.
For real time monitoring, one could generate a ruleset that identifies key metrics and calls the loan if metric limits are hit. In particular, metrics that should trigger loan recalls should include:
The benefits of this particular model is one can compare these probabilities in a tractible way. Rulesets defined by changes in probability make discussions and changes of lending parameters measurable leaving little room for interpretation. Lending protocols should have some method of default probability included, and can make adjustment based on the Merton Risk model as a starting point. More robust estimation could include substituting the collateral volatility with VAR which measures worst case losses, or taking half of the short term risk free rate as a return parameter.
Novel developments in the auction and recapitalization of bad loans are on-going in the crypto lending market space. Newer protocols have attempted continuous loans where collateral is auctioned off in real time to reflect price changes, effectively fixing a loan to value. These protocols engage in various flavors between continuous and partitioned auctions, but this model effectively is the generation of “risk tranches”. In traditional finance, risk tranches have lots of historic precidence where a user’s assets may be contractually sold off in chunks if price targest are hit, but a closed form model is just to take partitions of the users assets, and assert specific LTV requirements for each tranche.
While tranched lending isn’t particularly complex, one could imagine for a system with a new type of collateral and a target lending cap, to partition the cap into “chunks” with ranked LTV, where higher LTV commands higher payment rates. As capacity is reached on lower LTV (with cheaper rates), it is assumed the market would reach an equilibrium where most of the capacity with higher LTV is un-utilized. This particular system of course requires maintenance of lending rates and creation of new buckets as some risk tranches may be auctioned under high collateral stress, but serves to help protect lenders from a shock and major loss in a single lending agreement.
Since the cryptocurrency world has access and information to individual accounts, scoring debtors with credit risk metrics is possible. Some factors could include account volume traded, account age, maximum net worth, and other metrics that can help understand the risk profile of the lending agreement while taking into consideration the specific debtor. This could also be coupled with automated machine learning systems to develop account-specific lending agreements. While the decentralized finance space has yet to explore this area, traditional finance’s use of machine learning in credit risk models, taking into account variables is not new and regularly used. It may be prudent for mature lending markets to develop an in house quantitative risk team to profile user accounts. Cost-permitting, development of AI/ML tools to generate automated risk models using user-based features would be an entirely new product class, but could be worth the investment to help reduce losses for cryptocurrency based lending protocols, and possibly be a revenue generating product for other market participants.
Liquidity models of specific DeFi markets are not yet mature and mainstream. In particular, liquidity is linked to volatility and VAR since available liquidity soaks up asset sales and purchases, with large volumes trading in automated market making venues such as Uniswap or Curve. Issues arise with this automation, as decentralized finance operators are aware of Just-In-Time liquidity bots that serve to siphon MEV. Another factor are yield farming operations where protocols give additional value to market makers for providing token liquidity, giving rise to “mercinary capital” which may easily pull liquiditiy before market events occur. Because of these factors, it is unknown in general at any given time how much liquidity is available if a collateral shock occurs since these market making operations, whether automated or just in time, are not obligated to be available during market downturns.
While lending operators can take into account liquidity available on decentralized exchanges, it should serve no surprise that under a market shock, the available market makers can and will pull their liquidity from the market. The understanding of liquidity has a direct effect on volatility as it would be captured in returns as long as samples cover market shocks, which may lead lending practitioners to be wary of available liqudiity and volatility metrics when considering lending agreements. Also, it may benefit lending protocols to also engage in market making, and provide rewards and other equity returns to participants to soak downward pricing in collateral auctions, effectively engaging in a risk swap to offset potential losses.
Risk assessments in crypto credit markets is a relatively new field with many hurdles to overcome. While the space does have high volatility, there is still much to be considered as the blockchain space is a data rich environment. This document lighted VAR/Volatility, the Merton Default Probability Model, and Risk tranches as possible options for credit risk assessments that ultimately determine lending rates. While these should be considered as a first step, jhere is no shortage of new metrics to add to a panel of risk indicators, and lots of opportunities.
Along with this and the advent of the blockchain and AI systems, there are deeper risk modeling considerations that may be worth exploring for more well capitalized teams focused on assessing credit risk, and building quantitative risk systems. Prudent lending protocol users interested in risk frameworks should consider these basic options as a start. Operators engaging in DeFi lending markets should take these as baseline requirements at the start, and work to engineer their own in house automated risk operations.