Why blockchain CLOBs won't work

Centralized Limit Order Books (CLOBs) have become common on low-cost, low-latency chains such as Solana, EOS, Avalanche subnets, and Cardano. These new exchanges purport to solve many of the problems in Automated Market Makers (AMMs) like Uniswap, all while maintaining decentralization, cryptographic security, pseudonymity, and transparency. More specifically, these new exchanges say they have eliminated Miner Extractable Value (MEV), Impermanent Loss (IL), and slippage. As CLOBs are the standard in traditional finance, this would allow people to do what they have become accustomed to doing in trading stocks and making trades like limit and stop-loss orders. Given many have used Uniswap on Ethereum's mainnet, where fees are $5 to $50 per transaction and block times are 12 seconds, these new chains with fees under a penny and block times around a second would seem to allow CLOBs to work.

There are several reasons why decentralized CLOBs are a terrible idea. The first set applies to their claims about MEV, IL, and slippage; the last relates to the relative attractiveness relative to Bitmex, which is not on-chain but close enough for most people.

Trade Impact

In an AMM, the liquidity providers (LPs) are passive. They post tokens, and then any trade in the price range they are active shares the trade effect on a pro-rata basis. To save on capital, AMMs allow LPs to post tokens in restricted ranges, but this merely creates a set of step functions where this occurs; when the price is outside an LP's range, they are effectively irrelevant, receiving no reward and taking no risk. We can ignore restricted ranges without loss of generality to see the price impact mechanism CLOB advocates highlight.

Given a pool with ETH*USDC = constant, if a trader adds ETH to get USDC, we can solve for the USDC coming out:

(ETH + ethChange) * (USDC + usdcChange) = constant

The beautiful thing about this approach is everything flows out of this simple primitive. If we define the price as USDC/ETH, it can be shown that a trade will impact the price by

Price Change = 2 * usdcChange * sqrt(price/constant)

While this does not hold for tiny and huge trades for technical reasons, the point is that most trades will have a positive price impact, aka slippage.

In contrast, if we look at a CLOB, the bids are offers to buy specific quantities at specific prices. For example, in the snip below, one could buy 113.618 AVAX for exactly 3992.536 USDC, i.e., a set price of $35.140.

Actual CLOB for Avax-USDC on Dexalot

As with the case of Trader Joe's zero-trade-impact AMM, the CLOB has a zero-trade impact for small trade, as, on average, trading will have a trade impact. AQR is a large TradFi fund complex, and they estimated trade impact using their actual fill prices compared to initial prices and found it to be 15 basis points (bps). This sort of analysis can only be done using such data because if you merely look at tick data, it is impossible to link trades split up from a single trade. As many stocks have relatively small sizes offered at the best bid and offer, buying 10k shares in Apple stock with one trade is impossible, as the top of the book is usually only 200 shares. Traders at large shops use algorithms that split their trades into pieces, often using limit orders that look like a market maker (MM) order. Thus, high-frequency trading involves many orders and cancellations. One sees the result only at the end of the desired trade quantity, which might take a week.

In the 2010s, I worked at an equity-options trading firm designing algorithms for exiting delta positions acquired buying and selling options. It took millions of trades to get good data because the general fluctuations in the market create a lot of noise, and often, a sample of one year is dominated by a broad market trend that would, naively, imply all buys or sells had a negative trade impact. In my next position as a portfolio manager at Pine River, I took it upon myself to analyze the equity trades in our $6B multi-strat hedge fund, as we had many millions of shares traded, and I was curious. Some desks sent direct orders, but many were sent to algo trading desks at Goldman or JPMorgan. It generated the same pattern, all trading with a CLOB.

Trading out of a position will always cost you more than half the spread. Sure, if all you want to do is trade in and out at a size never larger than the top of the book, and you benchmark yourself to the mid, you know exactly what price you will pay—no trade impact—but that's irrelevant. It does not apply to any non-trivial position change and applied to very small trades, it loses money, so it's not like trading at zero-impact underlies any attractive trading strategy.

For both a CLOB and an AMM, sending a large market order relative to the liquidity generates a significant price impact. In both cases, the best thing to do is to split the trade up so as not to move the price too much. If you break up a trade of 1000 shares into ten 100 share trades, each trade may not have an impact, but in total, it does.

To the extent MEV exists, it is unavoidable and insignificant

MEV refers to the unique nature of blockchains, where sequencers have great latitude in jumping ahead of large trades. It is trivial to do on the Ethereum mainnet, where blocks are 12 seconds long. On CLOBs, this is done by HF traders and occurs in the order of milliseconds. The effects of these tactics are vastly overstated.

Reading Michael Lewis's Flash Boys gives the impression that HF traders are ripping off retail equity traders and making things worse. This ignores that trading costs have declined significantly since the introduction of electronic trading. Trade costs declined 7-fold from 1990 to 2010, while election trading rose from near-zero to 80%. In the good old days, you didn't have Virtu, Susquehanna, or Citadel running high-speed algorithms effectively making markets, but you did have decades of stasis where a small set of specialists were given a monopoly over order flow. Spreads on Microsoft were ¼ through the mid-90s, and its volatility was comparable to today. Currently, the spread is a penny while the price has risen 100-fold.

The fact that HF traders fight each other for a fraction of a cent highlights the perpetual criticism of competition, i.e., markets, in that whether it is Andrew Carnegie lowering the price of steel by 80% in the 19th century or Rockefeller enacting various efficiencies that cut the price of oil by 60% from 1860 to 1875, critics imagine how much better the world would be if such actors were eliminated. It's a pathetic ingratitude driven by envy and ignorance.

As per AMMs, the situation is similar. For anyone sending an order that is not outsized relative to the liquidity in the contract, there is no incentive to front-run the trade on chain because there is not enough value there to make it a problem. In practice, for most liquid pools, MEV is practically irrelevant.

MEV is problematic for the shiatcoins and illiquid pools tied to prices on other AMMs. However, the comparable here is not the NYSE but the pink sheets where penny stocks are traded. If you are trading a meme-coin or NFT, you are playing a zero-sum game at best, hoping to be on the up-side of a classic pump-and-dump. Avoiding MEV trading such coins is the third or fourth most pressing risk in the set of deceptive tactics afflicting your strategy, implying an exchange addressing this problem will have little value.

CLOBs replace IL with adverse selection

Market making in TradFi is very competitive. The high-frequency traders that act as electronic market makers do not have an explicit monopoly as in the specialist day. Instead, their barrier to entry is the time and money spent on algorithms and computers to make the right trades first. To compete, one needs colocated servers that respond within 50 milliseconds, if not 15 millis. Like most things, there is a power law distribution of ability, so there are a handful or dozens of such trading outfits, and all the other market makers, especially those trading through conventional trading GUIs sending limit orders at the best bid and offer, lose money.

For example, when working on high-frequency algos, I noticed that those orders at the top of the best bid or offer would make money, and those at the bottom would lose money. How do you get to the top of a queue? A speed advantage of milliseconds. This is not just about speed but having a better model, as jumping on a new temporary bid would also be a loser. Developing such algorithms by looking at patterns in the tick data takes a tremendous amount of intelligence and data, just as chatGPT needs billions of data points to be helpful. That is a barrier to entry that generates profits.

theoretical CLOB

The MM losers of this game suffer adverse selection. There are many scenarios where low latency is costly on CLOBs, but one applied to a market maker providing resting limit orders should suffice. A market maker places a two-sided order to buy or sell 100 shares of XYZ stock trading at a bid-ask price of $20.17-$20.18, its current bid price, $20.17. Assume the stock will move up or down $0.05 before you can cancel that order. If you get filled, it will only be because it is now trading at $20.12-$20.13, meaning you bought at $20.17 and can sell now at $20.12; you lost $0.05; if the price went up $0.05, you would probably not be filled on your $20.17 order to buy. Adverse selection in market making is when conditional upon getting filled, you paid too much or sold too low. It generates a loss profile for market makers like those selling straddles, a liquidity provider's impermanent loss, or a short straddle position.

In TradFi CLOBs, the best MMs can make money via their speed and better algos. Entry is open; virtually anyone can rent space for colocated servers. The best MM on blockchain CLOBs will probably also make money, unlike any of the LPs on an AMM. However, if you do not know you have privileged access to the order flow—the sequencer, RPCs, etc.—you do not. No independent third party is auditing the tape of NewDex on EOS, Injdojo on Injective, or Genius Yield on Cardano, which is like leaving a dog alone in the house with a cooked steak on the dinner table. Just as MEV is ineradicable on the Ethereum mainchain, the MM speed and access game is also unavoidable, whether acknowledged or not.

No price discovery for liquid coins

The most liquid coins trade on centralized TradFi exchanges where participants operate at under 50 millisecond speed. Any decentralized chain will have an order of magnitude greater latency because it requires consensus across a large geographic area to avoid centralization. The speed of light restricts the ability of such protocols from ever being comparable in speed to TradFi exchanges with colocated servers.

If you know the bitcoin price is increasing and buy on Binance and a slower blockchain CLOB, the price impact will show up on Binance first. The initial electronic exchanges like Island and Archapelego had greater price granularity and speed than the NYSE and Nasdaq, so they overtook these big institutions in setting prices. The only coins where blockchain exchanges do set prices are the smaller coins not listed on centralized exchanges, but these are not the coins emphasized by these new blockchain CLOBs.

Dexalot, a CLOB operating on an Avalanche subnet, acknowledged this as it introduced its CLOB by showing potential market makers how to set up a trading bot based on centralized CLOB prices.

The Bitmex advantage

Bitmex is a dependable crypto CLOB. It offers limit and stop orders and gives users 100x leverage. Sure, their perp financing rate is a scam benefiting their insider market makers, who almost surely have privileged access to order flow and a look at the stops of customer book, but where else can someone trade like that without KYC (one needs to use a VPN, but that's a minor annoyance)? They have proved themselves above the chicanery of Michael Lewis's idol SBF and have been through several market cycles. Bitmex has a good business and does not want to blow it up by gambling with customer money.

Further, one merely needs to deposit bitcoin to trade on Bitmex. They do not have a token, which is a friction for most protocols. The use case for these tokens is weak and focuses protocol development on the wrong priorities. For example, DyDx, FTX, and Serum are or were CLOB crypto exchanges that made their insiders billions based on fake volume generated directly or indirectly by exchange insiders. In all these cases, the token’s governance role and providing liquidity was or is lame. Convoluted scams attract bad people, so insiders tend to be knaves (SBF) or fools (Trabucco). Such protocols have a lot of operational risk. Bitmex is centralized, but it’s the best way for a kid to day-trade crypto and rightfully dominates the blockchain CLOBs. Bitmex trades about $400MM per day, almost entirely in perps where users can short and get leverage, about 100x greater than Dex CLOB volume.

DyDx paid out a net $110MM to pump its token to a $1.2MM market cap, their best trade

Conclusion

Blockchains need better exchanges. Developed countries' debt/GDP ratio has risen consistently from 60% to 100% over the past 13 years, and Hauser's law limits how much taxes can stop this trend. The endgame is ever-increasing inflation. Add to this that China and the developed Western countries are increasingly interested in tracking our every purchase for the ‘greater good,’ and decentralized crypto is our best hope for protecting liberty. The best on-chain exchanges—TraderJoe, Uniswap—are unsustainable due to their LPs losing money consistently. CLOBs might be a viable solution, but if so, they must accept they cannot deliver on the central promises they are making today and implement a mechanism to explicitly delineate the privileges of market makers.