Hyperliquid Blockchain: Microstructure & Perpetuals

• Hyperliquid is a Layer-1 blockchain with a fully on-chain limit order book for perpetual futures and spot trading, enabling detailed study of market microstructure and funding dynamics.
• It integrates HyperCore for order-book trading and HyperEVM for smart contracts, facilitating research on liquidity pooling, autodeleveraging, and risk management strategies.
• The platform's public-state execution data and visible TWAP orders offer actionable insights into execution costs, adverse selection, and capital efficiency under stress.

Hyperliquid is a purpose-built Layer-1 blockchain with a fully on-chain central limit order book, HyperCore, for spot and perpetual futures, plus a general-purpose smart-contract layer, HyperEVM. In recent research, it is treated primarily as a perpetuals venue and as the venue associated with the Hyperliquid liquidity pool, HLP: a setting in which protocol-level execution transparency, funding transfers, pooled liquidity, and venue-wide solvency mechanisms can all be studied from public state. The resulting literature analyzes Hyperliquid simultaneously as a market microstructure laboratory, a capital-efficiency mechanism for perpetual futures liquidity, and a case study in autodeleveraging under stress (Barone et al., 14 Jun 2026, Chitra et al., 9 Feb 2025).

1. Protocol architecture and public-state observability

Hyperliquid is described as a fully on-chain limit order book for cryptocurrency perpetual futures and spot trading. The venue combines HyperCore, which hosts the order-book exchange, with HyperEVM, a general-purpose smart-contract layer. Because all activity is on chain, researchers can reconstruct order flow, fills, and wallet-level participation from public data, rather than relying on proprietary audit trails or exchange disclosures (Barone et al., 14 Jun 2026).

This public-state property is central to the way Hyperliquid appears in the literature. One empirical study uses all 201 perpetual markets on Hyperliquid from July 28, 2025 to March 23, 2026, covering more than 641 million fills, about 365 million market orders, and roughly USD 1.93 trillion in notional volume. The same study constructs a separate order-book snapshot sample from December 15, 2025 to March 23, 2026 at one-minute frequency, with about 18,701 pair-day panels and 26.9 million snapshots. Because Hyperliquid trades continuously 24/7, order size and volatility are normalized using a rolling 24-hour window anchored at each order’s start time rather than calendar days.

The broader significance of this design is methodological. Hyperliquid provides a venue in which the market microstructure of perpetuals can be observed at protocol resolution, while also exposing data needed for studies of funding dynamics, liquidity provision, and solvency-restoration mechanisms. At the same time, later sections of the literature emphasize that this transparency at the trading layer coexists with more limited transparency in some aspects of pool management. This suggests a dual character: unusually transparent execution data alongside comparatively opaque liquidity-management logic.

2. Perpetual futures, funding, and the Hyperliquid liquidity pool

In the literature on decentralized perpetuals, Hyperliquid’s HLP is formalized as an instance of a Perpetual Demand Lending Pool (PDLP). A PDLP is a pool that lends assets to traders only for the purpose of trading perpetual futures on the associated exchange. The defining mechanics are that LPs deposit assets into the pool, traders borrow from the pool to open long or short perpetual positions, traders pay the pool a fee for the loan, positions are liquidated as soon as they become undercollateralized, and arbitrageurs and LP incentives help the pool maintain a target asset mix (Chitra et al., 9 Feb 2025).

Hyperliquid is placed in this category alongside GMX and Jupiter. The pool is described as permissionless on the LP side, but managed via a closed-source strategy by a single whitelisted entity, which adjusts the target portfolio and lends to traders. The same source states that Hyperliquid was a large exchange with over \$3 billion of open interest in December 2024. The paper does not provide Hyperliquid-specific formulas for exact HLP target weights, exact fee schedules, exact borrowing cost formulas, exact internal pricing rules, exact incentive parameters, or the exact rebalance algorithm. Hyperliquid therefore serves as a canonical live PDLP whose implementation details are not fully transparent.

The formal PDLP framework used to analyze Hyperliquid introduces reserves $R \in \mathbb{R}_+^n$, prices $p \in \mathbb{R}_{++}^n$, loans $\{c_i\}$, and a target weight. The pool weight vector is

$w(p,R)=\frac{p\odot R}{p^TR},$

and the pool seeks to keep $w(p,R)$ close to a target $w^\star$ by solving

$$\begin{aligned} &\text{minimize} && \left\Vert w(R+\Delta)-w^\star\right\Vert,\ &\text{s.t.} && \Delta\ge -R^A, \end{aligned}$$

where $R^A=R-\sum_i c_i$ is the available, unutilized portion of the pool. The solvency constraint is

$\sum_{i=1}^M c_i \le R.$

For perpetual funding, the paper uses the stylized linear model

$$\gamma_L(L,S,p,p_0)=\kappa\left(\frac{L}{S}-\frac{p}{p_0}\right),$$

where $p \in \mathbb{R}_{++}^n$0 and $p \in \mathbb{R}_{++}^n$1 are cumulative long and short open interest, $p \in \mathbb{R}_{++}^n$2 is the current underlying price, and $p \in \mathbb{R}_{++}^n$3 is the reference mark price.

A central claim of the PDLP analysis is that such pools are often easier to delta hedge than CFMM positions. The main delta-hedging solution is

$p \in \mathbb{R}_{++}^n$4

with the zero-transaction-cost form

$p \in \mathbb{R}_{++}^n$5

The interpretation given is that the optimal hedge offsets the pool’s risky exposure $p \in \mathbb{R}_{++}^n$6, while fee income from lending supports taking that hedge. In this framework, Hyperliquid’s significance is not a publicly specified HLP formula, but its role as a large real-world PDLP whose scale helps motivate the claim that pooled lending to perpetual traders can be capital-efficient and hedgeable.

3. Visible execution, TWAPs, and sunshine trading

Hyperliquid is also studied as a rare empirical testbed for sunshine trading theory. The venue offers protocol-native TWAP orders whose parent order is visible from inception and remains visible while active, together with ordinary order flow from which hidden metaorders can only be inferred ex post. The protocol automatically splits a large TWAP into child orders every 30 seconds, aims for a roughly time-proportional schedule, enforces a maximum slippage constraint of 3% on each slice, and allows later slices to catch up if earlier ones underfill, with catch-up capped at three times normal slice size (Barone et al., 14 Jun 2026).

Using address-level data, researchers reconstruct about 4.3 million statistical metaorders and about 465,000 native TWAP metaorders after restricting attention to reconstructed metaorders with at least 10 child orders, native TWAPs with at least 5 child orders, and executions completed within 24 hours. The execution schedule is formalized as

$p \in \mathbb{R}_{++}^n$7

where $p \in \mathbb{R}_{++}^n$8 is the cumulative fraction executed by normalized time $p \in \mathbb{R}_{++}^n$9. Uniform execution corresponds to $\{c_i\}$0.

The two execution regimes differ sharply. Hidden statistical metaorders are front-loaded and U-shaped, consistent with transient-impact optimal execution and risk aversion. Native TWAPs are much closer to a uniform schedule. The impact functions also differ in shape: native TWAPs are roughly power-law in traded fraction $\{c_i\}$1, with estimated $\{c_i\}$2, much flatter than the classic square-root benchmark, while statistical metaorders are better fit by a curved arctangent form. The paper emphasizes that the square-root law is not universal in this setting.

The main empirical findings concern execution cost and adverse selection. Visible TWAPs have lower temporary impact than comparable hidden metaorders; in the common-support surface comparison, the median log-difference corresponds to a TWAP cost advantage of about 2.3x. In pooled regressions, the TWAP indicator coefficient is about $\{c_i\}$3, translated by the authors into roughly 8.9 basis points lower temporary impact for a visible TWAP relative to a statistical metaorder at median volatility. Native TWAPs also leave a smaller permanent price displacement: the TWAP coefficient is around $\{c_i\}$4 to $\{c_i\}$5 bps in the baseline specification, and around $\{c_i\}$6 to $\{c_i\}$7 bps even without the traded-fraction control.

The nonannouncer result is equally important. A 10 percentage point increase in same-side visible TWAP dominance raises the permanent impact of hidden metaorders by about 0.84–0.92 bps, and after controlling for the mechanical price pressure of overlapping visible flow, a residual same-side effect of about 6 bps remains. During active native TWAP windows, oriented imbalance rises, displayed depth increases, sweep costs fall, and the inside spread widens slightly. In the event-time regression, the active-window coefficients are about $\{c_i\}$8 bps for relative spread, $\{c_i\}$9 for imbalance, $w(p,R)=\frac{p\odot R}{p^TR},$0 USD for depth, and $w(p,R)=\frac{p\odot R}{p^TR},$1 bps for sweep cost. Larger TWAPs elicit more imbalance and more depth.

The literature interprets these results as support for Admati–Pfleiderer-style sunshine trading: preannounced execution on Hyperliquid does not mainly trigger predation, but tends to attract liquidity provision and improve execution quality for the announcer, while shifting some adverse-selection costs toward hidden same-side traders.

4. Funding-aware market making on Hyperliquid perpetuals

A separate line of work treats Hyperliquid as a perpetual DEX in which inventory management is inseparable from funding-carry exposure. A market maker on Hyperliquid manages not only mark-to-market inventory risk, but also the fact that a long contract can generate funding income or funding expense depending on the sign of the funding rate, with the opposite sign for shorts. Funding is therefore modeled as a stochastic state variable rather than a static fee (Le, 7 May 2026).

The paper converts Hyperliquid’s reported fractional funding rate $w(p,R)=\frac{p\odot R}{p^TR},$2 into a cash-scaled funding state,

$w(p,R)=\frac{p\odot R}{p^TR},$3

so that funding enters the control problem in the same units as spread capture and PnL. The state is $w(p,R)=\frac{p\odot R}{p^TR},$4, where $w(p,R)=\frac{p\odot R}{p^TR},$5 is cash, $w(p,R)=\frac{p\odot R}{p^TR},$6 inventory, $w(p,R)=\frac{p\odot R}{p^TR},$7 mark price, and $w(p,R)=\frac{p\odot R}{p^TR},$8 cash-scaled funding per unit inventory. Inventory evolves as

$w(p,R)=\frac{p\odot R}{p^TR},$9

with quote-dependent fill intensities

$w(p,R)$0

Funding is modeled in the baseline as an Ornstein–Uhlenbeck process,

$w(p,R)$1

and cash evolves as

$w(p,R)$2

The reduced HJB contains the term $w(p,R)$3, which is the theoretical expression of the fact that if $w(p,R)$4 and $w(p,R)$5 have the same sign, funding is costly, while if they have opposite signs, funding is beneficial. Optimal quotes are recovered from neighboring inventory value differences, yielding

$w(p,R)$6

with a quote floor applied afterward.

Calibration uses Hyperliquid ETH, BTC, and SOL perpetuals, hourly funding observations, one-minute L2 mid-price panels, and official crossed fill data. The funding process is fit first with a Gaussian OU baseline and then checked against an OU-plus-jump diagnostic. The estimated funding half-lives are about 5.560 h for ETH, 4.071 h for BTC, and 2.310 h for SOL. The reported hourly jump probabilities are 2.05% for ETH, 1.24% for BTC, and 0.16% for SOL, with OU-plus-jump log-likelihood gains of 2816.34, 4417.03, and 7149.78 respectively. The interpretation is explicit: Gaussian OU is retained because it is tractable and monotone-discretizable, but the data strongly suggest that jump funding is a better future extension.

Empirical evaluation is a 100-seed holdout simulation on 26 Nov 2025 to 31 Dec 2025 under two official-fill proxy calibrations, volume_minute and minute_hit. Under volume_minute, ETH shows final equity of 74471.93 for pure_as and 75784.42 for hjb_fd, a delta of +1312.49, a win rate of 0.62, and inventory RMS of 3.6475 versus 5.7362. BTC shows 46808.28 for pure_as and 47617.37 for hjb_fd, a delta of +809.09, a win rate of 0.57, and inventory RMS of 0.1831 versus 0.2974. SOL shows 85989.45 for pure_as and 104860.37 for hjb_fd, a delta of +18870.92 and a win rate of 1.00, but inventory RMS of 51.8427 versus 20.0173. Under minute_hit, ETH and BTC remain positive relative to classical AS, while SOL again gains with much larger inventory exposure.

The research conclusion is narrow but important: on Hyperliquid, funding is a relevant state variable for liquidity provision, and a funding-aware HJB can improve mean ETH/BTC performance while lowering inventory RMS relative to classical Avellaneda–Stoikov. SOL is presented as the cautionary case in which higher profits do not constitute a Pareto improvement once risk scaling is accounted for.

5. Autodeleveraging, solvency restoration, and the October 10, 2025 stress episode

Hyperliquid is also the main empirical case in two formal studies of autodeleveraging (ADL), the last-resort loss socialization mechanism used when liquidations and insurance buffers are insufficient to restore solvency. In the online-learning formulation, Hyperliquid-style ADL operates on a PNL-haircut domain: haircuts apply to positive unrealized PNL, not to posted collateral principal. Exchange solvency is written as

$w(p,R)$7

with residual shortfall

$w(p,R)$8

and winner haircut capacity

$w(p,R)$9

Each ADL round chooses a solvency budget and an allocation over winners from the feasible set

$w^\star$0

The dynamic-severity regret bound is

$w^\star$1

The paper further shows that queue allocations are extreme points of the feasible polytope, are not Lipschitz or continuous, and that the unique optimizer of the min-max fairness problem is pro-rata (Chitra et al., 16 Feb 2026).

Applied to the Hyperliquid stress episode of October 10, 2025, this study examines the interval 21:16–21:27 UTC, with \$w^\star$2100.1M, an aggregate needed budget of approximately \$w^\star$360.1M. Production overshoot is reported as \$w^\star$445.0M–\$w^\star$564,859,522.21 at $w^\star$6, versus \$w^\star$74,413,367.61 for vector mirror descent. The instance-calibrated upper envelope is approximately \$129.7M; production is about 50.0% of this bound, while the best start-of-round deployable baseline is about 2.6%.

A second ADL study uses the same Hyperliquid event to develop a broader impossibility and optimization framework. It describes Hyperliquid’s production mechanism as a queue-based PNL-leverage ranking rule descended from early BitMEX/Huobi-style ADL logic, and proves a trilemma: no static ADL policy can simultaneously achieve exchange solvency, revenue preservation, and fairness to traders. In that model, budget balance is

$w^\star$8

with policy outputs consisting of a severity $w^\star$9 and a haircut vector $$\begin{aligned} &\text{minimize} && \left\Vert w(R+\Delta)-w^\star\right\Vert,\ &\text{s.t.} && \Delta\ge -R^A, \end{aligned}$$0. The paper proposes capped pro-rata, risk-aware pro-rata, and dynamic or Stackelberg ADL controllers, and argues that queue-based ADL is worst for the top winner (Chitra, 30 Nov 2025).

For the October 10 event, this second study reports a 12-minute window in which ADL was used repeatedly to close \$\begin{aligned} &\text{minimize} && \left\Vert w(R+\Delta)-w^\star\right\Vert,\ &\text{s.t.} && \Delta\ge -R^A, \end{aligned}$1304.5M, feasible haircut capacity of about \$\begin{aligned} &\text{minimize} && \left\Vert w(R+\Delta)-w^\star\right\Vert,\ &\text{s.t.} && \Delta\ge -R^A, \end{aligned}$2208.6M, production queue budget use of about \$\begin{aligned} &\text{minimize} && \left\Vert w(R+\Delta)-w^\star\right\Vert,\ &\text{s.t.} && \Delta\ge -R^A, \end{aligned}$3630.5M, residual loss after queue of about \$\begin{aligned} &\text{minimize} && \left\Vert w(R+\Delta)-w^\star\right\Vert,\ &\text{s.t.} && \Delta\ge -R^A, \end{aligned}$447.2M. On this basis, it states that Hyperliquid’s production queue overutilized ADL by approximately $\begin{aligned} &\text{minimize} && \left\Vert w(R+\Delta)-w^\star\right\Vert,\ &\text{s.t.} && \Delta\ge -R^A, \end{aligned}$5 relative to the paper’s optimal policy and imposed roughly \$630 million of unnecessary haircuts on winning traders.

The numerical gap between the two ADL studies is not necessarily a contradiction. One adopts a PNL-haircut accounting with an aggregate needed budget of approximately \$\begin{aligned} &\text{minimize} && \left\Vert w(R+\Delta)-w^\star\right\Vert,\ &\text{s.t.} && \Delta\ge -R^A, \end{aligned}$660.1M, while the other evaluates queue budget use against aggregate deficits and feasible haircut capacity. This suggests that the estimates are benchmark-dependent rather than directly interchangeable. What the two studies share is the conclusion that fixed queue logic is structurally weak under repeated stress, and that separating severity from allocation, together with proportional or adaptive allocation rules, materially improves tracking and fairness.

6. Limitations, open problems, and research directions

The current literature repeatedly emphasizes that Hyperliquid is simultaneously unusually observable and only partially transparent. The trading layer is public enough to support address-level reconstruction of order flow, fills, wallet participation, and even ADL execution behavior, but the HLP liquidity pool is described as using a closed-source strategy managed by a single whitelisted entity. As a result, formal PDLP results are not a direct reverse-engineering of Hyperliquid’s internal logic (Chitra et al., 9 Feb 2025).

Several analyses are explicitly stylized. The PDLP work uses simplified linear funding and abstract target-weight models, while noting that real protocols may have more complex pricing, liquidation, oracle, and risk-management rules. It also leaves multi-period arbitrage for future work and proposes dynamic parametrization of fees, dynamic target portfolios, more formal treatment of hedged PDLP strategies, and possibly splitting pools or tailoring them to covariance structure in order to improve capital efficiency. In the market-making work, Gaussian OU funding is retained because it is tractable and monotone-discretizable, but OU-plus-jump diagnostics indicate that a future nonlocal jump-HJB would be more realistic (Le, 7 May 2026).

The market-microstructure evidence also comes with identification caveats. Some active liquidity providers may be misclassified as metaorder executors; native TWAP choice is highly endogenous to trader identity; statistical metaorders are not perfectly anonymous because all addresses are visible; and the hidden benchmark is therefore a latent but partially learnable regime rather than true anonymity. The same study does not find a strong cross-market relation between the TWAP discount and market-level adverse-selection proxies such as Kyle’s lambda, price-impact $$\begin{aligned} &\text{minimize} && \left\Vert w(R+\Delta)-w^\star\right\Vert,\ &\text{s.t.} && \Delta\ge -R^A, \end{aligned}$$7, and order-flow autocorrelation, suggesting that sunshine-trading effects appear more clearly at the conditional event level than in coarse cross-market comparisons (Barone et al., 14 Jun 2026).

ADL research identifies a separate set of open problems. Dynamic severity control, proportional or risk-aware allocation, and replay-based instance-calibrated bounds are proposed as practical design improvements, but the broader trilemma implies that no mechanism can fully optimize solvency, fairness, and revenue simultaneously. This gives Hyperliquid a distinctive place in current research: it is not only a live perpetuals venue, but also a public laboratory for studying how on-chain transparency, pooled liquidity, funding-carry dynamics, and solvency-restoration rules interact in a modern cryptocurrency derivatives market (Chitra et al., 16 Feb 2026, Chitra, 30 Nov 2025).