The two most popular mechanisms for onchain trading are AMMs and Dutch auctions. Both leak MEV to arbitrageurs.
For AMMs, this is now well understood as loss vs rebalancing (LVR).
What about Dutch auctions?
New paper with
@danrobinson
🧵
Thrilled to have
@ciamac
join
@paradigm
as a Research Advisor!
@ciamac
is a Columbia finance professor and quant who is one of the sharpest and most practical researchers I've met
He coauthored the LVR papers, which set the gold standard for influential academic DeFi research
New paper alert!
We propose a novel AMM design that seeks to address two interrelated questions (1) how should an AMM set the trading fee? (2) How can an AMM mitigate LVR / liquidity provided losses to arbs.
Interesting discussion between
@0x94305
and
@AndreaCanidio
, both of whom are experts and have thought deeply about DEX design, on evaluating the real-world performance of
@CoWSwap
AMM.
Two questions I'd like to discuss: 🧵
@0x94305
I think to evaluate a new type of AMM, you need to check whether LPs make more money by LPing in the new AMM than the alternative. The rest is an interesting intellectual exercise.
We are thrilled to announce a $7M raise to become the leading Open-Access AI Cloud 🤘🏼🌪️
At Hyperbolic, we’re building an open AI ecosystem and economy where everyone who contributes is rewarded. Our goal is not to merely optimize AI performance to compete with traditional Web2
It's a sad time to be a
@Columbia
faculty member.
Abandoning principles of academic free expression, our administration has decided to double down on censorship and McCarthyism under a pretext of "safety".
@NadiaAbuElHaj1
explains:
Longer block times lead to less CEX/DEX MEV per second but more atomic MEV per second.
I’m aware that the profit formulae imply otherwise, but looking at just the math misses the sequential nature of the availability of information.
1/5
New version of our paper on the economics of liquidity provision in automated market marking, clearer conceptual presentation along with some empirical results 👇🏼
So
@jason_of_cs
,
@ciamac
,
@Tim_Roughgarden
and I posted an update to our paper on automated market makers and loss-versus-rebalancing! Here's the abstract and link!
Diving deeper into data gives us the bigger picture.
Tomorrow, our data researchers,
@xin__wan
&
@austinadams10
will sit down with Columbia Professor
@ciamac
to discuss is work at Columbia, dark pools, impermanent loss, and more 👀
🚀 Optimal Dynamic Fees for Blockchain Resources: Introducing a new framework for designing dynamic fee mechanisms for multiple network resources. (Codename: Endgame 1559)
Joint work with
@ciamac
and
@Qiaoqiao2001
from the DRO division at Columbia GSB.
TL;DR & link in 🧵👇
Why not show users loss-versus-rebalancing? Fees-LVR accurately estimates (delta-hedged) liquidity provider P&L. It's actually not difficult to account for.
3/ Today, historic price, volume, and token info are scattered across different sources, and it's difficult to account for liquidity depth and divergence loss.
That changes today.
Inspired by a trading terminal, the Liquidity Terminal places data LPs need at their fingertips.
800K doses delivered, 337K used, leaves 463K doses remaining. Of these, 253K are needed for dose 2. That leaves 210K remaining. NYC is doing 25K/day of dose 1.That leaves about 8 days of supply. OTOH If they are proactively holding back dose 2, that leaves about 4 days of supply.
2/ The goal is to understand the impact of fees on arbitrage trading against AMMs, and use this as quantitative guidance to understand how to set fees and how to design AMMs to minimize the MEV extracted by arbs and tradeoffs therein.
13/ This is consistent with the observations of many (e.g.,
@0x94305
@MaxResnick1
) that faster blocks are an easy way to mitigate DEX MEV, perhaps at the cost of reducing decentralization.
5/ Under the assumption of Poisson block generation, our first result is to solve for the steady state distribution of the DEX-CEX mispricing, which follows a jump diffusion process. This allows us to quantify the probability of the no-trade region.
This is an interesting question! My intuition is that arb profits/LVR in an AMM can be viewed as a (liquidity weighted) continuously sampled realized variance. Some geometric intuition for that below.
Let me try to settle the question of block times and profits of CEX-DEX arbitrageurs. Arbs' profits are LOWER but LESS volatile with shorter block times. LPs' losses to arbs thus are also lower but more certain when blocks are produced faster... [1/n]
6/ We show that, if fees are gamma and mean interblock time is Delta t, the probability that a block contains a trade (the probability of being outside the no trade region) takes a simple form:
3/ The starting point is LVR, i.e., how much do DEX LPs lose to DEX-CEX arb in an idealized setting, trading in continuous time (no discrete blocks) and with no trading fees
@jason_of_cs
@Tim_Roughgarden
and myself are doing a virtual tutorial on "Automated Market Makers and Loss-versus-Rebalancing (LVR): a Deep Dive into Decentralized Finance", as part of Preview Week for
@AcmSIGecom
EC 24
@lpachter
@lexfridman
And how can he interview Joscha Bach
@Plinz
with no mention or discussion of his long term funding and support from pedophile Jeffrey Epstein?
4/ What happens when we incorporate discrete block generation and trading fees? Both are frictions that impact arbitrage trade. Fees create a “no-trade” region, where although the DEX and CEX prices differ, the mispricing does not exceed the fee and hence arbs don’t trade.
@paulg
Yes, you are following noise. PredictIt prices didn't even satisfy basic arbitrage relationships. There are too many fees/restrictions and basically no professionals in this market, pure noise trading.
1/ I promised a while ago a thread on how the losses of Uniswap LPs to arbs depend on fees, volatility and block times. But then got distracted - first with markouts, then Curve stablecoin, then FBAs, and then a Twitter space on AMMs.
Time to return to the topic eternal...
@MarketUrbanism
Tourists are a big problem in NYC, IMO. They impose congestion and other externalities. Maybe fewer tourists would be better, the city should optimize for it's residents.
🦄 Introducing
In collaboration with the good folks at
@Uniswap
, we've built a dashboard powered by Uniswap v3 shadow events
Showcasing an example of the types of data-rich apps that are easy to build with Shadow
17/ One way to think about the choice of fee is through the framing of
@rithvikra0
and
@theshah39
: fees create a tradeoff between losing money to arbs and the accuracy of the pool prices.
The second talk from the Digital Finance Seminar Series
@Columbia_Biz
has been posted.
Watch
@ObadiaAlex
and
@0xQuintus
from Flashbots talk about "Why You Should Care About Maximal Extractable Value (MEV)":
7/ This probability depends on the fee measured in units of typical return (stdev) over half the interblock time. When fees are high or the interblock time is low, it becomes less likely that arbs can profit on any given block. For example:
Apparently too few citations in the work of
@timnitGebru
is a fireable offense at
@GoogleAI
. Why was she not allowed to see the review? Why not just admit she was fired? This statement makes no sense because it is PRed+lawyered nonsense.
I understand the concern over Timnit’s resignation from Google. She’s done a great deal to move the field forward with her research. I wanted to share the email I sent to Google Research and some thoughts on our research process.
11/ Note that there is an interesting discontinuity here: when fees are zero, arb profits are basically LVR — they do not vary much with the interblock time.
15/ Though LVR was developed assuming no fees and continuous trading, even with fees and discrete blocks, LVR is roughly the profit gross of fees of arbing the pool. Introducing fees simply changes how LVR is split and who earns it (arbs or pool LPs).
9/ The formulas simplify when fees are low and blocks are frequent (the “fast block” regime), in this case arb profits are simply LVR scaled down by the probability of trade.
We show how practitioners can trade off speed and quality of execution by parameterizing the auction’s starting price and decay rate based on the block time, the asset’s volatility, and uncertainty about the asset’s fair value, determining an efficient frontier.
I also discuss the pros and cons of different practical mitigation strategies being developed:
1. faster blocks
2. dynamic AMMs
3. ex post auctions
4. ex ante auctions
Thus, the seller should expect to sell the asset at a discount, with the profits going to MEV.
We call this loss vs fair (LVF). It is analogous to LVR in AMMs.
@jason_of_cs
@Tim_Roughgarden
@alz_zyd_
SUMMARY: There are some nuances and design choices in doing a markout analysis, but markouts (or some other delta-hedged P&L analysis) are crucial in evaluating any trading mechanism!
END
I give an overview of the high level theory behind DeFi MEV, namely loss-versus-rebalancing (LVR) for AMMs and for Dutch Auctions, and how it relates to block times and multi-block MEV.
We build on earlier work quantifying arb profits / LP losses for AMMs with Poisson block generation and with asset prices following geometric Brownian motion.
For example, if volatility σ=5% (daily), decay rate δ=1bp/sec, and average block time Δt=12 sec, LVF is about 13bp. This would mean that for every $100 worth of tokens that they sell, the seller should expect to get about $99.87.
In an onchain Dutch auction, the true market price of the asset at the time a block is created may be higher than the price offered by the auction, due to the decay of the Dutch auction price and the drift and volatility of the asset.
Enjoyed
@thelatestindefi
? Want to see what algorithmic game theorists think about DeFi? Submit a paper to the EC24 DeFi workshop!
We'll have some fun keynotes from
@Tim_Roughgarden
and
@ciamac
But bear in mind when you start looking at unhedged measures of LP profitability, you are introducing market risk into your calculations, and this is extremely noisy and less likely to be predictive of future outcomes.
@benedictevans
FB is not under any obligation to sue these specific academics in this instance. They are choosing to do so and the most likely explanation is not that they suddenly care about privacy.
Our results illustrate how characteristics of the blockchain — particularly the average block time — impact efficiency. To support Dutch auctions that lose less than 2bp for assets with daily volatility of 5%, block times of less than 2.75 seconds are needed.
As
@0x94305
observes, and this is closely related to P&L after delta-hedging: LVR isolates profit and loss from participating in the DEX from profit and loss because token price went up (or down).
The solution is to try multiple offsets and check that the results are robust. We did this in the LVR paper, in a delta-hedged P&L analysis of UniV2 WETH-USDC, and found that the time offset/delta-hedging interval didn't matter so much from 4 hours down to 1 minute.
We also consider the expected time-to-fill (for regular Dutch auctions) and the expected sales rate (for continuous Dutch auctions), and derive closed-form expressions for these.
Q1. Are markouts the right tool for evaluating LP profitability in a novel DEX such as CoW AMM?
Q2. How to define the reference/fair price (on-chain prices vs. binance prices, whether and what time offset to use, etc.)?
Q1: Absolutely yes! Markouts are very intuitive and get at the LVR for *any* trading mechanism. If you are LPing in a DEX (AMM, LOB, or whatever), a very natural question is are you trading at "fair" prices?
@ConorMcMenamin9
@0xEzon
@danrobinson
I'm not sure this answers the question, but if there are no fees, LVR (per unit time) is (to a first order) independent of the block time. If there are fees, LVR scales with sqrt(block time), and faster chains experience less LVR.
We can quibble about the exact nature of the markout (and I will in a moment), but I think ***some type of direct LVR analysis*** (be it markouts, or delta-hedged P&L) is fundamental and IMHO first line of analysis in looking at a DEX.
Crafting the Crypto Economy is hosted by
@CBER_Forum
members Fahad Saleh, a Professor of Finance at Wake Forest University, and Andreas Park (
@financeutm
), a Professor of Finance at the University of Toronto.
@bertcmiller
Indeed a fun challenge! How about this: draw a graph with bundles as vertices, and an edge between bundles that share a transaction. Each "permutation" is an independent set:
I had a great chat with super-smart
@ciamac
on quant investing, the pros and cons of machine learning and how to think about trading frameworks from time horizons to costs. Oh and we talk bitcoin too :)
Q2: For the markout price, it's best to use a reference price with the most liquidity and smallest spreads. For WETH/USDC, Binance prices are better than using on-chain UniV3 prices, but I agree with
@0x94305
that this difference is not likely to be large.
This is the heart of loss-vs-rebalancing (LVR): the "rebalancing" strategy does the same trades as fair prices, and hence LVR is conceptually a comparison between actual trades done and hypothetical trades at fair prices.
The first talk from our Digital Finance Seminar Series
@Columbia_Biz
is up!
Watch
@ebudish
talk about "The Economic Limits of Bitcoin and Anonymous, Decentralized Trust on the Blockchain":
The model in our paper is written as if the arb hedges on the CEX after each DEX trade. That said, I suspect it's optimal not to hedge on the CEX per trade at all. Instead, the CEX price is viewed as a "signal" for when to trade on the DEX and how much.
Three Columbia students sued the school in housing court saying they were illegally locked out of their dorms after being arrested or issued a citation at pro-Palestine protests
Story by
@ramseykhalifeh
for
@Gothamist
-
@WNYC