Perpetual futures under the hood

Last, index and mark: three prices on one screen

A perpetual (a perp) is a futures contract that never expires. A dated future is tied to spot because on its settlement date it becomes spot. A perp has no such date, so the venue needs two extra prices to keep it honest, and every serious perp screen shows all three.

The mark price is where the design choices live. Binance takes the median of the index adjusted for pending funding, the index plus a moving average of the basis, and the contract’s own price. Hyperliquid takes the median of three inputs: the oracle plus a 150 second moving average of its own premium over the oracle; the median of its best bid, best ask and last trade; and a median of perp prices on large centralized exchanges.

A median of several inputs means one thin print cannot move it, so a single wick or spoof on one venue does not wipe out positions. The flip side: you can be liquidated while the last price on your chart never touched your level, and survive a wick that went straight through it. How far a venue can trust its inputs is a question about oracles.

Funding: the tether that replaces expiry

When the perp trades above the index, longs pay shorts. When it trades below, shorts pay longs. The payment makes the crowded side more expensive to hold and pays anyone willing to take the other side, which pulls the contract back toward the index. As Binance puts it, funding is transferred directly between traders holding opposing positions, and it charges no fee on it. You only pay or receive if you hold a position at the funding timestamp.

The standard formula, used with small variations by Binance, Bybit, Hyperliquid and others, has two parts:

The premium index measures how far the contract trades from the index, using the price a meaningful order would actually get rather than the top of the book:

On Binance that fixed notional, the impact margin notional, is 200 USDT divided by the initial margin rate at maximum leverage: 4,000 USDT for a contract with 20x maximum leverage. dYdX uses 500 USDC divided by the initial margin fraction. Sizing it this way makes the premium expensive to fake with a few small orders.

The interest component represents the difference between borrowing dollars and borrowing the coin. On Binance it is fixed at 0.03% a day, which is 0.01% per 8 hour interval, for most symbols. The clamp then does something subtle: whenever the premium sits within 0.05% of the interest rate, the two cancel and funding is exactly the interest rate. That is why 0.0100% is the most common number you will ever see on a funding screen.

The payment is position notional times the rate, with notional valued at mark on Binance and at the spot oracle price on Hyperliquid. The third row is what happens with an index of 60,000, an impact bid of 60,090 and an impact ask of 60,110: a 0.15% premium becomes 0.10% funding, and a 60,000 dollar long pays 60 dollars for the interval. Every venue also caps the rate. Binance sets the cap and floor at 0.75 times the maintenance margin ratio, Hyperliquid caps funding at 4% an hour, and dYdX caps the 8 hour rate at 600% of the gap between initial and maintenance margin.

Funding intervals differ by venue

The formula looks alike everywhere, but the clock does not. The details below are as of September 2026. Comparing a raw funding number across venues without normalizing the interval is the most common mistake in reading funding data.

To compare, convert everything to one interval and then to a yearly rate:

An hourly 0.0013% is about 0.0104% per 8 hours, the same as Binance at 0.0100%. An hourly 0.01% is 87.6% a year. Reading derivatives data covers what those levels say about positioning.

Initial margin, maintenance margin, cross and isolated

Initial margin is what you post to open: notional divided by leverage. Maintenance margin is the floor your equity must stay above; cross it and the liquidation engine takes over. On Hyperliquid, maintenance is half the initial margin at the market’s maximum leverage, so a market with 40x maximum leverage (2.5% initial) has a 1.25% maintenance rate, and across its markets the rate runs from 1.25% to 16.7%. Centralized exchanges use tiers instead: the maintenance rate climbs as the position grows, because a larger position is harder to close without moving the book.

For an isolated position, set equity equal to the maintenance requirement and solve for price:

Take a 1 BTC long at 60,000 with 10x isolated leverage, so 6,000 dollars of margin, and m = 1.25%. The liquidation price is 54,683.54, a fall of 8.86%, not 10%. At 54,000 the equity would be exactly zero: that is the bankruptcy price. The 683.54 dollars between the two is the maintenance margin at that price, and it exists to pay for closing your position in a falling market without anyone else taking the loss. The same trade as a short liquidates at 65,185.19, 8.64% above entry.

# Isolated margin liquidation price, before fees and funding.
# side = +1 for a long, -1 for a short. m is the maintenance margin rate.
def liq_price(entry, leverage, m, side=+1):
    return entry * (1 - side / leverage) / (1 - side * m)

liq_price(60_000, 10, 0.0125, +1)   # 54,683.54  (long, 8.86% below entry)
liq_price(60_000, 10, 0.0125, -1)   # 65,185.19  (short, 8.64% above entry)

Liquidation, insurance funds and auto deleveraging

A liquidation is a sequence of backstops, each one used only when the one before it runs out. The details below are Hyperliquid’s, because its documentation is public and specific; centralized venues follow the same shape with an insurance fund in the middle.

1. The mark price pushes equity below maintenance

   Nothing has been sold yet, and there is rarely any warning beyond the margin ratio creeping up.

2. The engine closes into the book

   Hyperliquid first sends market orders to its own order book. Positions above 100,000 USDC are closed 20% at a time, with a 30 second cooldown between orders. If the closes restore the maintenance requirement, any remaining collateral stays with the trader.

3. A backstop takes the position

   If equity falls below two thirds of maintenance before the book can absorb it, a backstop liquidation hands the position to the liquidator vault and the maintenance margin is not returned. HLP, a community funded protocol vault that also makes markets, performs these liquidations. Centralized exchanges use an insurance fund instead: it keeps what is left when a liquidation fills above the bankruptcy price and pays the shortfall when one fills below it.

4. Auto deleveraging, the last resort

   If an account or isolated position goes negative and the backstops cannot cover it, the venue closes profitable positions on the other side. Hyperliquid ranks them by (mark price ÷ entry price) × (notional ÷ account value), so the most profitable and most leveraged go first, and closes them at the previous mark price against the underwater account.

October 10, 2025 showed the whole stack under load. After a US tariff announcement, more than 19 billion dollars of leveraged positions were liquidated within 24 hours, and total perp open interest fell from about 217 billion to 123 billion dollars, according to CoinDesk research; Hyperliquid’s open interest fell 57%. An academic analysis of auto deleveraging counted about 2.1 billion dollars of Hyperliquid positions closed by ADL in 12 minutes that day.

Onchain perps: order books versus oracle pools

Onchain perpetuals come in two families, and the difference is who stands on the other side of your trade.

Order book venues

Hyperliquid keeps an order book for every asset in its own chain state, matched in price time priority, with margin checked when an order is placed and again when it fills. dYdX runs as its own proof of stake chain built on the Cosmos SDK, where validators hold the order book in memory, off chain and outside consensus, and put matched trades into blocks. In both, your counterparty is another trader or a market maker, as on a centralized exchange. See order books, RFQ, intents and solvers for how books match.

Oracle priced pools

GMX takes the opposite approach. There is no book: every order is routed against liquidity pools and quoted at an oracle index price from Chainlink Data Streams. The pool’s liquidity providers are the counterparty to every trader: they earn a share of the fees and lose when traders win. Instead of premium based funding, GMX v2 uses adaptive funding that rises while one side dominates open interest, a borrowing fee on the larger side, and a price impact charge that rewards balancing the pool. Positions close automatically once collateral net of losses and fees falls below 0.4% to 1% of size, depending on the market, and ADL closes profitable positions if pending profits threaten the pool.

The pool design makes the pool’s own accounting part of the attack surface. On July 9, 2025, GMX v1 on Arbitrum lost about 42 million dollars to a reentrancy bug: the list of short positions was updated but the global average short price was not, which inflated the value of the GLP pool token so the attacker could redeem it for more than it was worth. The attacker returned about 40.5 million dollars after a bounty offer.

A position from open to close

Put the pieces together on one illustrative trade: a 1 BTC isolated long at 60,000 with 10x leverage, a 1.25% maintenance rate and a taker fee of 0.05% each way. Funding is charged on the entry notional for simplicity; real venues use mark or oracle notional at each interval.

The fourth row is the one people forget. The price never moved, yet a 4.3% dip now liquidates the position instead of an 8.9% one, because isolated margin pays the funding. You can run this experiment, with funding charged every bar, in the trading simulator, which offers spot, perpetuals and dated futures with play money.

Five checks before you open a perp

1. Find the mark price recipe

   Which inputs, which venues, what smoothing. That is the price that liquidates you, and on a thin market it is only as good as its index.

2. Compute your own liquidation price

   Use the formula with the venue’s real maintenance tier, then subtract a few weeks of funding at a pessimistic rate. The liquidation calculator and position size calculator help.

3. Normalize the funding rate

   Convert to the same interval and annualize before comparing venues or deciding whether a carry is worth holding.

4. Choose cross or isolated deliberately

   Isolated for a trade you are prepared to lose outright, cross only when you are watching the whole account.

5. Read the backstop rules

   Insurance fund size, backstop vault, ADL ranking. They decide what happens to you when someone else’s position fails.

The risk management page covers sizing so that a liquidation is an event you planned for. Primary sources worth reading in full: Binance funding rates, Hyperliquid funding, Hyperliquid liquidations and the GMX documentation.

Common questions

Where to go next