Perp DEX market design

00 · AbstractIt's all about the flow

Wholesale market makers paid Robinhood roughly a billion dollars in 2024 for the privilege to service its retail order flow; Citadel led the list, with Virtu, Susquehanna, and Two Sigma close behind. Similar payments flowed from those firms to Schwab, eToro, Webull, and the other consumer brokerages. The same dynamic has played out with searchers and block builders over the past few years on EVM and SVM platforms. Enter the perps DEX era, where hundreds of exchanges are competing for the same sort of retail flows.

We see a variety of misguided approaches. Stipends paid for posting size produce books that look real at rest and evaporate when they are needed. Sub-millisecond latency races decay participation to a handful of incumbents and gating the heterogeneous trader cohort which comprises healthy markets. Points programs underwriting mercenary capital that rotates the moment the incentives dry up. The same instruments trading everywhere with increasingly fragmented liquidity and inefficient pricing.

In this article we'll be going down the rabbit hole on market design parameters for these venues, exploring what it takes to create healthy markets.

01 · ValueWhat is flow worth, anyway?

The mechanics of payment for order flow are well known to some. Consumer brokerages route their retail order flow to wholesale market makers in exchange for cash payments. Robinhood's 2024 disclosure put the cash payments at roughly $1.06 billion across all venues, of which Citadel was the largest single payer at an estimated three-to-four hundred million.

These arrangements persist because all parties benefit. The wholesaler earns the spread on a population of trades that does not systematically know more than they do. The retail customer typically receives price improvement and/or a no fee trade. The broker collects the per-trade payment and uses it to subsidise the user-facing product.

In contra, consider a market made entirely of informed participants trading at each other. Every quote a free option to whoever has the better information. The participant with the worst information gets adversely selected first and leaves. The participant with the next-worst information degenerates to the lowest man on the totem pole, gets adversely selected, and leaves. A venue that succeeds at attracting professional makers and fails at attracting retail takers becomes a thin book of professional makers picking each other off, and that book eventually closes. Market design is, at its root, a retail flow acquisition, engagement and retention problem.

02 · MeasuringThe quality of order flow

This begs the question of how to measure the quality of flow. For example, a venue with $1B of daily volume could be largely retail flow, or stipend receiving makers rebalancing inventory against each other, or points farming wash traders.

One measure of quality is order flow toxicity, where toxicity is basically the relative information content of the flow from some viewpoint. VPIN or Volume-Synchronised Probability of Informed trading¹ a descendant of the PIN model² one of the best known ways to look at this. VPIN flagged the 2010 Flash Crash roughly an hour before it happened,³ and a Lawrence Berkeley National Laboratory study commissioned by the SEC called it "the strongest early warning signal known to us at this time."⁴

To apply VPIN we partition trading into $n$ buckets, each of size $V$ in notional. Within bucket $i$, let $V_i^B$ and $V_i^S$ be volume classified as buyer-initiated and seller-initiated, with $V_i^B+V_i^S=V$. Then

VPIN uses Bulk Volume Classification (BVC),⁵, which runs on bars and splits each bar's volume fractionally between buy and sell using the standard normal CDF of the standardised price change between consecutive bars,

where $Z$ is the standard normal CDF, $V_{\tau }$ is volume in bar $\tau$, and $\Delta p_{\tau }$ is the price change between bars $\tau -1$ and $\tau$.

A bucket with balanced buy and sell volume basically means the participants on either side were uninformed enough that the bucket cleared without one side picking off the other. A lopsided bucket means the dominant side probably had information. Persistently elevated VPIN is the population-level signature of toxic flow. Makers facing it widen, then withdraw, and a feedback loop pulls the remaining liquidity out.

Another way to look at this is the with Shannon entropy of resting orders and matches:

with $p_i$ the fraction of resting volume at price level $i$. On FX data,⁶ quote-distribution entropy moves with fast-trader share and dampens the price impact of macro news, with a 10% bump in entropy reducing the adverse impact of news on effective spreads by roughly 60%. The same mechanism applies to perp DEXes: a richer quote heterogenicity indicates a healthier market. The more entropy settled and resting orders contain, the more noise there is in the book.

In lieu of deeper analysis, there's a more heuristic tell: daily volume divided by open interest. Liquid futures at TradFi venues run 0.5x to 3x daily-volume-to-OI. A venue running say a 10:1 volume to OI ratio, or higher, is mostly recycling the same notional through the book without anyone taking directional positions, which is what wash-trading for points looks like.

The lower the toxicity of order flow, or the higher it's entropy, the more desirable it is to service. This is a key performance indicator venues should pay attention to, and that market makers use to value out integrations.

The best flow is generally on venues like Hyperliquid and Binance, where price discovery is actually occurring. The lion's share of perp DEXes don't have a PFOF channel; with no broker layer segregating retail from informed flow, everyone's orders land in the same lit CLOB. Most of that flow fails the measures above, which is why the majority of venues have to pay makers to show up. How much subsidy a venue has to pay is a function of how toxic its taker side actually is.

03 · DifferentiatingWhat brings traders?

The Cold Start Problem⁷ frames any two-sided marketplace bootstrap as a chicken-and-egg problem. In the bootstrapping phase you need to subsidise experience for both sides until the network hits a density at which each side's presence is itself a sufficient reason for the other side to show up. The book's worked examples are consumer products, but the principle generalises. For perp DEXes the wrinkle is saturation, there's no exponentially better venue coming.

Retail flow is the be-all-end-all in a healthy network of market participants. But, before retail can come, liquidity must exist. The strategy we've seen work around this over the past few years has been to build a profitable trading strategy, or partner with a firm who has one, and run that as a vault product. HLP is the canonical example: an in-house liquidity pool which serviced the early taker flow, kept the spread-capture economics inside the venue, and pulled in user capital as a yield product. Lighter took a different route with variable delay on taker order processing plus a vault trading against the delayed flow, which economically replicates payment-for-order-flow and pays the captured spread back to depositors as yield.

Hyperliquid had more than one advantage. Jeff Yan and the founding team came out of a quantitative trading background, so the API, the order types, and the product semantics were designed by someone who'd seen the good, the bad, and the ugly. The HLP vault strategy came out of years of experience with Chameleon. The crypto and SV cohort founding many other perp DEXes learned the mental model second-hand. Dog fooding your product is crucial, and being your own biggest user is a great way to do that. These advantages come with offsetting fragilities in the current implementation: no US access without a wrapper, validator-set centralization that is only now expanding, and HLP concentration risk. Of course, no incumbent is immune to disruption.

Places venues can differentiate live in the ecosystem around exchanges. A first-class PFOF channel that segregates retail from informed flow at the brokerage layer doesn't exist on any perp DEX as a product, and prime-brokerage-level capital efficiency that beats CEX cross-margin only just coming to fruition. Instrument selection beyond the majors is where Hyperliquid has gone furthest of late with HIP-3⁸ opening permissionless deployment of new perp markets, and HIP-4⁹ extending the same concept into outcome markets.

04 · Fee dialsCarrots and sticks

Most perp DEXes pay makers via a flat monthly stipend tied to a posting requirement, a per-fill rebate funded out of taker fees, or a points-based deferred allocation.

A stipend contract might look like: post $X within Y bps of mid for Z% of the trading day, get $W per month. These arrangements result in the maker posting the minimum size at the widest allowable spread for the minimum required uptime, and pulling during the windows the venue most needs liquidity. Inversely, per fill rebates incentivize the maker to keep quoting through adverse selection and volatility. Rebates are then funded by taker fees. Most venues with maker rebate tiers pair them with taker fees. Interestingly taker fees by themselves increase flow quality by debouncing, if you will, micro pips of the book.

Market makers who move meaningful volume often work additional private terms: a multiple of the published rebate, a guaranteed minimum that approximates a stipend, latency improvements, custom fee schedules, sometimes equity in the exchange. This has to be managed carefully, as some market participants being more equal than others can produce strange and unexpected failure modes in market formation.

Then of course there are points programs: trade now, earn points, and at some point in the future those points convert to tokens at a rate the platform decides.

This is unsecured credit at zero interest with non-deterministic settlement. The market participant fronts the fee outflow today with no contract specifying how many points convert to one token, when the TGE happens, what the token's launch price will be, or whether the program will be diluted by future seasons. The exchange takes the fee revenue today and pays it back, maybe, in tokens whose supply it controls. As a financing arrangement for the venue, that's excellent. As a compensation mechanism for the participant, it's a bet on the venue's TGE economics.

The capital that shows up for the points is mercenary by construction. It's there to extract a token allocation, and the moment a more attractive program launches at another venue, the capital rotates. Points work fine as a supplement. A venue with a desirable product can layer points on top to fund growth, recognise contributors, and pre-distribute the token to the participants who actually built the venue. A venue using points as user rent has markets that exists for only as long as the points last.

05 · LatencyTime is money

Intra venue latency varies widely. Hyperliquid takes 250ms to 500ms to get an order into the book, Lighter is roughly 250ms with deliberate delay for takers, venues like hibachi are sub 10ms. A new firm with a solid stack and well constructed infrastructure can participate without going full FPGA. With that said, perps DEXes have begun to aggregate into ap-northeast-1 in amazon, and latencies are pushing downwards.

Whether latency is a constraint introduced by venue implementation, which it largely is at present, or by the laws of physics, thoughtful design is a key differentiator for venues. Hyperliquid is one of the slowest perps dexes at present, and thus leaves makers exposed to adverse selection for hundreds of milliseconds. To combat this they sequence each block as add liquidity only orders, or ALOs post first, then cancels, then aggressors, so makers get defensive priority on both posting new liquidity and pulling stale quotes. Lighter takes a different route: variable delay on order ingestion for aggressing orders.

A venue with deliberate latency design has a wider revenue surface. A venue that doesn't design around latency loses. The maker leaves because their rebate doesn't cover the markouts. The retail taker leaves because their fills are systematically worse than CEX.

06 · Price discoveryFair value

Price discovery generally happens on the highest volume and lowest latency venue for a given market. For ETH, for example that's Binance, or for BTC during US business hours, that's CME futures. Price discovery is a distributed process, and as soon as some instrument trades across venues each venue's book contains some information about the price, with latency dynamics shaping how quickly that new information propagates.

If you look at each price series individually, it's non-stationary, but the linear combination of price series on instruments based the same underlying assets is stationary, or to say it in another way, cointegrated.¹⁰ One model for working with cointegrated time series is a vector error-correction model, or a VECM, which treats the these correlated prices as a single system with a common long-run efficient price plus short-run deviations that revert back toward it. Once you fit a VECM to a basket of venues, you can ask which one's innovations are doing the most work moving the common price, and that's what tells you who's leading.

There are some common measures which attribute that leadership in different ways. Hasbrouck Information Share¹¹ decomposes the innovation variance of the common efficient price across venues. Gonzalo-Granger Permanent-Transitory decomposition¹² reads the VECM's error-correction coefficients to identify which venue does less of adjusting back to equilibrium, and therefore is driving the permanent component. Hayashi-Yoshida lead-lag¹³ is a tick-by-tick cross-correlation built for asynchronously sampled trade data; the lag at which venues' returns are most correlated tells you the lead-lag time.

Run on ETH spot across five high-volatility events in 2024, all three measures converged on Binance leading Uniswap v2.¹⁴ Binance's Hasbrouck share ran 0.96 to 0.99, the Hayashi-Yoshida lead-lag ratio ran 1.11 to 1.86, and a separate cross-correlation analysis on the same pair has Binance returns predicting Uniswap returns 20 minutes ahead with no detectable signal in reverse.¹⁵ BTC tells the same story compressed with CME futures leading Binance spot.

A perp DEX trading the same underlying as Binance is structurally downstream in price discovery. This is because the institutional flow, credit infrastructure, and product breadth concentrate at Binance, and a fast matching engine alone on a competing DEX doesn't change where any of that lives. The same logic applies why we see CME lead binance during US hours for BTC, combined of course with a seasonality component as the world turns.

The opportunity for DEXes here is to list instruments that don't trade upstream. Hyperliquid has built furthest in this direction, with HIP-3 builder-deployed perps and HIP-4 outcome markets putting HL on the price-discovery side for assets that didn't previously exist on a CEX. Polymarket holds the same position for prediction markets, Kalshi for events.

07 · ConclusionWhat wins

Hyperliquid is the DEX which has won thus far. The industry shape will eventually resemble most others: two or three big winners, a tail of ten or so large venues. All of the winners will utilize the best practices walked through above. Namely, per fill rebates, deliberate latency structure that protects makers, vaults that service early flow, unique and desired instrument selection and structure.

In terms of the greenfield opportunities, cross-venue margin is one. A user trading across Hyperliquid, edgeX, Lighter, and Aster shouldn't have to fund a separate margin account at each. The TradFi answer is the FCM: holds the user's collateral, posts margin to CME, ICE, LME on the user's behalf, runs unified risk. Cross-asset margin within a single venue is solved by venues like Binance with unified margin, and Hyperliquid with spot-perp cross margin.

A broker that manages flow across venues another. Route the user's orders to whichever venue prices best, hold idle collateral as yield-bearing balance, capture the spread between user execution and venue execution, rebate a share back to the user as fee credits or yield. This is best-ex and PFOF in one product. The TradFi version paid Robinhood roughly a billion dollars in 2024. The on-chain version doesn't exist yet.

Other opportunities remain in PFOF channels to route segregated flows, RFQ rails to let large takers source firm prices at size, dark pools that cross block size at midpoint without leaking intent, and closed RFQ for bilateral OTC.

The next winners will combine the best practices that have become table stakes with high quality liquidity, enhanced capital efficiency and alternative routes into the market built to suit various species of users.