Funding rates and open interest

How to compute and annualize funding correctly, why venues differ, how the cash-and-carry trade really works, and what open interest can and cannot show.

Lesson 14 of 15Intermediateabout 30 minAll lessons

You’ll learn

  • Compute funding payments and annualize a rate correctly, and name the common traps
  • Explain why funding differs across venues and over a venue’s own history
  • Build the profit and loss of a cash-and-carry position, including costs, capital and failure modes
  • Define open interest and trace how it changes with new, transferred and closed positions
  • Separate what funding and open interest can support as inferences from what they cannot
Real story

When one exchange’s funding split from the rest

During the November 2022 collapse of FTX, then the fourth-largest crypto exchange, Bitcoin perpetuals on FTX traded substantially above spot, while on solvent exchanges they traded below it. Funding rates turned deeply negative on the solvent venues. He, Manela, Ross and von Wachter read this as consistent with FTX customers closing short futures positions quickly, either voluntarily to cut exposure to a failing exchange or involuntarily through liquidation. Identical contracts on one asset showed opposite funding because they had different counterparty risk. The number was recording who wanted out of what, in a specific place.

Source: <a href='https://arxiv.org/abs/2212.06888'>He, Manela, Ross and von Wachter, Fundamentals of Perpetual Futures, Section 2 and Figure 2</a>

Calculating funding

Lesson 11 derived the mechanism. For analysis, three facts matter: the rate is a percentage of notional for one interval, it is paid only by positions open at the funding time, and its size depends on venue-specific settings. The payment is simple.

Payment = q × Pref × F

q is signed position size, Pref the venue’s reference price for converting to notional, and F the rate. A positive result is paid by a long position. Binance uses the mark price. Hyperliquid states that it uses the oracle price.

A payment at a live reading

  1. Hyperliquid’s ETH oracle price read $2,388.76 and hourly funding 0.00125% on November 5, 2026.
  2. Example position: 3 ETH long. Notional = 3 × 2,388.76 = $7,166.28.
  3. Payment for the hour = 7,166.28 × 0.0000125 = $0.0896, paid to shorts.
The amount looks trivial per hour. It is charged on notional, though, and the next section shows what it means on an annual basis.
VenueIntervalInterest componentStated cap or limit
Binance USDⓈ-M8 hours by default; some contracts 1 or 4 hours, and it can change in extreme volatility0.01% per 8 hours by default; ETHBTC uses 0%±0.75 × maintenance margin ratio on listed symbols; ±2% on other USDⓈ-M contracts
HyperliquidHourly, one eighth of the computed 8-hour rate0.01% per 8 hours0.5% per hour
dYdX v4Hourly, premium ÷ 8 plus interest0% in cross markets; 0.125 basis points an hour in isolated markets8-hour rate capped at 600% × (initial − maintenance margin): 12%, 30% or 60% by tier
GMXAccrues over timeNo such term. The larger side pays the smaller, at a rate that adapts with imbalanceRate moves toward a per-market floor. A separate borrow fee also applies

As documented in November 2026. Hyperliquid notes that its cap is much less aggressive than its centralized counterparts.

Binance states that a trader pays or receives only if a position is open at the funding time, and that the actual time can deviate by up to 15 seconds. A position opened one minute after a funding time and closed before the next pays nothing, however large the rate. So funding is not a smooth cost on an 8-hour venue, and any calculation of carrying cost must be built from funding timestamps.

Annualizing and its traps

Traders annualize to compare funding across assets, venues and interest rates. There are two standard forms.

Simple: A = F × n    Compounded: A = (1 + F)n − 1

F is the rate per interval and n the number of intervals per year: 1,095 for 8-hour funding, 8,760 for hourly.

The same rate, two annual numbers

  1. 0.01% per 8 hours, simple: 0.0001 × 1,095 = 10.95%.
  2. 0.01% per 8 hours, compounded: (1.0001)1095 − 1 = 11.57%.
  3. 0.00125% per hour is the same rate. Simple gives 10.95%. Compounded gives 11.57%. Hyperliquid’s documentation states this as “11.6% APR”, which matches the compounded result.
  4. A rate of 0.1% per 8 hours gives 109.5% simple and 198.8% compounded.
Always state which form is used. At high rates the difference is large, and compounding assumes a rate persists and that proceeds can be reinvested, both of which are often false.

Each of the following is a separate way to misread an annualized rate.

Common mistakes

  • Extrapolating one print. A 109.5% rate from one 8-hour reading describes one interval. Funding can flip sign within days. Report an average over a stated window.
  • Comparing different bases. Funding is on notional. At 10x, a 10.95% annual rate on notional is 109.5% of initial margin. Neither number is wrong, but they answer different questions.
  • Reading the default as a measurement. The clamp produces the default rate over a wide band of premiums. An average near 10.95% can mainly reflect how often the clamp held.
  • Mixing intervals. A 0.01% rate per hour is eight times a 0.01% rate per 8 hours. Check a data vendor’s unit before comparing.
  • Ignoring compensation for risk. A rate can be high because the short side is risky to hold, not because money is free.
  • Assuming capture. A published rate is what a position holder would have been paid or charged, before spread, fees, and the risks listed below.

A short exercise on real data makes the clamp visible. The following script retrieves BTC hourly funding on Hyperliquid and pages through the API’s 500-row response limit. Without paging, a 30-day request silently returns about 21 days.

import json, time, urllib.request URL = "https://api.hyperliquid.xyz/info" def post(payload): req = urllib.request.Request(URL, json.dumps(payload).encode(), {"Content-Type": "application/json"}) with urllib.request.urlopen(req, timeout=30) as r: return json.load(r) def funding_history(coin, days): now = int(time.time() * 1000) start = now - days * 86_400_000 rows = [] while start < now: page = post({"type": "fundingHistory", "coin": coin, "startTime": start}) if not page: break rows.extend(page) start = page[-1]["time"] + 1 # the API returns at most 500 rows per call time.sleep(0.2) return rows rows = funding_history("BTC", 30) rates = [float(r["fundingRate"]) for r in rows] n = len(rates) simple = sum(rates) / n * 24 * 365 compound = 1.0 for r in rates: compound *= 1 + r compound = compound ** (24 * 365 / n) - 1 base = 0.0000125 print("hours:", n) print("simple annualized:", round(simple * 100, 2), "%") print("compounded annualized:", round(compound * 100, 2), "%") print("share of hours at the 0.00125% baseline:", round(sum(abs(r - base) < 1e-9 for r in rates) / n * 100, 1), "%") print("share of hours negative:", round(sum(r < 0 for r in rates) / n * 100, 1), "%") print("min, max hourly:", min(rates), max(rates))

Run on November 5, 2026, this returned 720 hours for BTC with a simple annualized average of 8.78%, a compounded figure of 9.18%, 70.0% of hours at exactly the baseline, 6.9% of hours negative, and hourly extremes of −0.00179% and +0.00125%. So for 30 days, the highest print was the baseline itself, and all the variation in this sample lay below it. That is a description of one window on one venue. It shows how little of the average carried any information beyond the formula’s default.

Why funding differs across venues

He et al. report that funding rates on BTC tend to be similar across exchanges because of cross-exchange arbitrage, and that they diverge during extreme liquidity episodes. They measure a correlation of 0.74 between Binance and Bybit, which is high but leaves meaningful divergence during periods of large payments. They identify two main forces: arbitrage is segmented, because exchanges do not offer margin across venues, so each leg needs its own collateral; and exchanges differ in liquidity, so the same order flow moves the price more on a thin venue. Binance accounted for approximately 49% of BTC perpetual volume among the five venues they study, for 2020 to 2022, and Bitfinex for 0.2%.

Added to those are differences by design: the formula, the interest component, the interval, the reference price, and whether the contract is linear or inverse. dYdX’s zero interest default on cross markets means its rate has no 0.01% baseline, so a raw comparison with Binance mixes market effects with design effects. And collateral matters. A venue’s rate partly reflects what trading there involves: stablecoin risk, the chance of a withdrawal halt, or exchange solvency. The FTX case is the extreme form. Part of a funding gap may be payment for that risk rather than a free spread.

The cash-and-carry basis trade

Definition

Cash-and-carry: Holding an asset in spot and shorting a futures or perpetual on the same asset in equal size, to earn the premium or funding while roughly neutralizing price exposure. Practitioners also call the perpetual version funding-rate arbitrage.

With a dated future, the return is locked at inception. An example: spot at $2,000 and a 90-day future at $2,030 is a 1.5% basis, or 6.08% annualized on a simple basis, if the position stays open to expiry and nothing breaks. With a perpetual there is no locked amount. The return is whatever funding turns out to be, and He et al. stress that the strategy is not risk-free even without margin or fees, because no predetermined expiry forces a profitable unwind.

Example: 30 days of a perpetual carry position

  1. Example inputs: $100,000 spot, and a $100,000 short perpetual with $20,000 of margin. Total capital is $120,000.
  2. Fees: spot 0.10% per side, perpetual 0.05% per side. Total = 2 × $100 + 2 × $50 = $300 for a round trip.
  3. Funding received: 0.01% per 8 hours for 90 intervals = 0.9% × $100,000 = $900.
  4. Net result if basis is unchanged: 900 − 300 = $600.
  5. Return on capital over 30 days: 600 / 120,000 = 0.5%, or 6.08% a year on a simple basis.
  6. If funding had averaged −0.005% per 8 hours instead: −$450 − $300 = −$750.
The headline rate of 10.95% became 6.08% once costs and capital were counted, before any other risk. Holding period and capital structure change the answer substantially.

The evidence is that this premium has been large and that it comes with real exposure. Schmeling, Schrimpf and Todorov find that carry on crypto futures averages above 10% a year and can reach 60% a year, far larger than in equities, bonds, currencies or commodities. They describe the trade as risky because of spikes in margins and liquidations during price drawdowns. He et al. report Sharpe ratios for a trading rule using their no-arbitrage bounds, for BTC, of 3.35 under high retail-tier costs and up to 11.65 with zero fees, on Binance data to March 2024. They also report maximum drawdowns no higher than 3.90% for BTC and up to 13.90% for DOGE across their specifications. These are backward-looking results from one venue, with their own assumptions, and deviations narrowed by about 22 percentage points a year over the sample. They do not show that any future return is available.

RiskMechanismReal example or evidence
Funding reversalFunding turns negative and the position paysHe et al. document negative funding in March 2020 and November 2022
Liquidation of the short legA price rise generates a margin loss on the short. Spot gains cannot post to the margin account in time without cross-collateralIn the example above, a rise of about 18.5% would hit the short’s liquidation price at 1.25% maintenance. Schmeling et al. note margin spikes
Auto-deleveragingThe short is closed by rule, leaving the spot unhedgedFTI Consulting reports this for hedged shorts on Oct. 10, 2025
Collateral failureA stablecoin or staked token used for margin falls in valueUSDe, BNSOL and wBETH on Binance, Oct. 10, 2025
Venue failureExchange insolvency or a withdrawal freezeFTX, November 2022
Basis widening before convergenceMark-to-market loss, possibly with forced closure before the gap narrowsHe et al. discuss interim losses and liquidation limits
Capacity and crowdingMany holders of the same trade compress returns, and may exit togetherDeviations shrink over time in He et al., and Schmeling et al. link arbitrage capital scarcity to carry

Not a complete list. Taxes, borrow cost if the spot is financed and operational failures are also real.

He et al.’s evidence on portfolio margining shows why capital structure is central. After Binance’s April 2022 change, an arbitrageur could net collateral across spot and futures, so less capital was tied up for the same position, and deviations compressed. The same netting can carry a cost, because it ties survival to how the venue values collateral in stress, as the October 2025 case showed for some collateral on one venue.

What open interest measures

Definition

Open interest (OI): The CFTC’s definition: the total of all futures and/or option contracts entered into and not yet offset by a transaction, by delivery, by exercise, etc. It also states that aggregate long open interest equals aggregate short open interest. Open interest is a stock, while volume is a flow.

TradeBuyer of the contractSeller of the contractChange in OI
1Opens long 100Opens short 100+100, to 100
2Opens long 50Opens short 50+50, to 150
3Opens long 40Closes long 40 (a sale)0, to 150. A position changed hands
4Closes short 30 (a purchase)Closes long 30 (a sale)−30, to 120

Example. Volume across the four trades is 220 contracts while OI ended at 120. Trade 3 is volume with no change in OI at all.

Several measurement issues follow directly.

Units. OI may be stated in contracts, coins or dollars. A dollar figure moves with price. In example numbers, 1,000 ETH of OI is $2,000,000 at $2,000 and $1,800,000 at $1,800, a 10% fall with no position closed. Comparing over time should use coin terms, or show both. On November 5, 2026, Hyperliquid’s API reported BTC OI of 38,522.9 BTC, about $3.07 billion at the mark price of $79,681, against $4.05 billion of 24-hour notional volume. It reported ETH OI of about 1,027,200 ETH, about $2.45 billion.

Counting convention. One contract with one long and one short counts as one in the CFTC’s convention. Chitra’s paper warns that some exchange reporting can differ by a factor of two from an unmatched-position definition. Before adding across venues, check each venue’s definition. Binance’s endpoint documentation does not state the unit of its open-interest field.

Coverage. A market-wide OI total covers only the venues in the aggregator. Vendors differ in that list, so CoinDesk Data’s −27.5% (nearly $60 billion to $158 billion) on October 10, 2025 and the “about $65 billion” calculation by Bitwise’s Jonathan Man are not necessarily in conflict.

Open interest cannot show direction. Every long has an equal short, so OI cannot rise because “longs are piling in” in any sense that excludes shorts. It shows how much is outstanding, and says nothing of who is on which side.

What can and cannot be inferred

ObservationWhat it supportsWhat it does not support
Funding well above the defaultThe contract has traded above the index, and the long side has paid to holdThat longs are right or wrong, or that price will fall or rise
Funding exactly at the defaultThe premium sat within the clamp bandThat the market is neutral. A wide band produces this
Rising OI with a rising priceNew positions are opening as price rises, on both sidesThat “new longs are in control”. A new short opened every one of those positions as well
Falling OI in a crashPositions closed, some possibly forcedThat all were forced. Liquidation totals are separate and have their own reporting limits
High OI relative to spot depthForced closure could have a large impact, under the cascade logic in Lesson 13That a cascade is coming, or at what level
Long-short ratio from a venueThat venue’s account or position split, as definedDollar balance. Accounts differ in size, and total long always equals total short by notional

Each row states the limit of a common reading. None of these is a timing tool.

The popular four-quadrant rule (rising price with rising OI is “strong”, rising price with falling OI is “short covering”, and so on) is a descriptive shorthand with little testing behind it. The “short covering” reading of a rising price with falling OI is reasonable as far as it goes, since a closing purchase by a short removes open interest when the seller is also closing. But the seller in that trade was closing a long, which the rule omits. Evidence for predictive value is a separate question, requiring the methods from Lesson 9.

Open question

Is high funding a measure of crowded leverage or of scarce arbitrage capital?

Both readings have support, and they are not mutually exclusive. Schmeling et al. link high carry to demand from small investors seeking leveraged upside and to limited arbitrage capital. He et al. find that past returns predict the deviation, consistent with extrapolative demand, and that price impact is larger when deviations are already wide, consistent with binding arbitrage capacity. Their portfolio-margin evidence points to the supply channel, since a change in capital efficiency was followed by compressed deviations. So funding is a joint outcome of demand and arbitrage supply, and a single rate cannot identify which moved. Treating high funding as proof of crowded longs alone assigns a cause that the number cannot carry. The direction of any link from funding to later returns is also unsettled, and depends on the sample, venue and horizon.

Common mistakes

  • Saying funding is “the cost of leverage” in general. Short positions can receive it, and the rate depends on premium rather than leverage used.
  • Adding long-side and short-side OI, then calling the total market size.
  • Comparing dollar OI across time while ignoring price change.
  • Quoting a carry return without capital, fees or basis change.
  • Comparing funding from a pool-based venue with a premium-based venue as if the two were the same instrument.
  • Drawing conclusions from one venue’s funding when the venue’s share of the market is small or has shifted.

Problem set

  1. An example 8-hour rate reads 0.03%. Compute simple and compounded annualized rates. Then compute the funding paid over 30 days by a long position of $50,000 if the rate holds.
    Show solution
    Simple: 0.0003 × 1,095 = 32.85%. Compounded: (1.0003)1095 − 1 = 38.9%. Over 30 days, 90 intervals × 0.03% × $50,000 = $1,350, or 2.7% of notional.
  2. Venue A pays hourly at 0.004% and venue B pays every 8 hours at 0.03%. On a simple annualized basis, which is higher? Give two reasons why an observed gap may not be capturable.
    Show solution
    A: 0.004% × 8,760 = 35.04%. B: 0.03% × 1,095 = 32.85%. A is higher by 2.19 percentage points. Reasons a gap may not be capturable include: trading fees and spread on two venues, separate margin on each leg, venue and collateral risk, rates that change before the position opens, and interval timing.
  3. Trace OI: Trader A opens long 200 against B opening short 200. Then A sells 120 to C, who opens long. Then B buys 80 from C, with C closing. What is the final OI, volume, and each trader’s position?
    Show solution
    After trade 1: OI = 200, A +200, B −200. After trade 2: A +80, C +120, B −200, OI = 200 because a position moved. After trade 3: B −120, C +40, A +80, OI = 120. Check: longs 80 + 40 = 120, shorts 120. Volume = 200 + 120 + 80 = 400.
  4. In the carry example, what average funding per 8 hours over 30 days would give a zero net result if fees are $300? What if capital earned 4% a year elsewhere?
    Show solution
    Zero net needs $300 of funding: 300 / (90 × 100,000) = 0.00333% per interval. To match 4% a year on $120,000 over 30 days, the position must also earn 120,000 × 0.04 × 30/365 = $394.52, so funding must total $694.52, or 0.00772% per interval.
  5. An example ETH OI goes from 500,000 ETH at $3,000 to 520,000 ETH at $2,400. What are the coin and dollar changes? Write a correct one-sentence description.
    Show solution
    Coin OI rose 4%. Dollar OI went from $1.5 billion to $1.248 billion, a fall of 16.8%. A correct sentence: open interest rose 4% in ETH terms while its dollar value fell 16.8%, because the price fell 20%. A statement that “traders closed positions” would be false on these numbers.
  6. A commentator says funding is at a default level, so “leverage has been flushed out.” Critique the claim using the clamp.
    Show solution
    A default print means only that the average premium was in the clamp band, here −0.04% to +0.06% per 8 hours. Leverage is a feature of account balances relative to notional, which funding does not measure. OI may still be high, and funding could be at default both before and after a flush. Evidence would need OI in coin terms, and account or margin data where available.

References

  1. He, S., Manela, A., Ross, O., and von Wachter, V. (2022, this draft September 2026). Fundamentals of Perpetual Futures. arXiv:2212.06888. Source
  2. Schmeling, M., Schrimpf, A., and Todorov, K. (2023). Crypto carry. BIS Working Paper 1087. Source
  3. Ackerer, D., Hugonnier, J., and Jermann, U. (2024). Perpetual Futures Pricing. NBER Working Paper 32936. Source
  4. US Commodity Futures Trading Commission. Commitments of Traders: Explanatory Notes. Source
  5. Binance. Introduction to Binance Futures Funding Rates (read November 2026). Source
  6. Binance. Open Interest endpoint, USDⓈ-M futures API (read November 2026). Source
  7. Hyperliquid. Funding documentation (read November 2026). Source
  8. Hyperliquid. Info endpoint, including pagination limits and perpetuals data (read November 2026). Source
  9. dYdX. Funding documentation (read November 2026). Source
  10. GMX. Trading fees documentation (read November 2026). Source
  11. Chitra, T. (2025; v3 February 2026). Autodeleveraging: Impossibilities and Optimization. arXiv:2512.01112. Source
  12. FTI Consulting (December 24, 2025). The Crypto Crash of October 2025: When Leverage Met Liquidity. Source
  13. CoinDesk Data (October 28, 2025). How CCIX Performed During Crypto’s Largest Liquidation Event. Source

Exercise

  1. Run the script in this lesson for BTC, then for a smaller asset. Compare the share of hours at baseline and the share of hours negative.
  2. Change the window to 7, 30 and 90 days. Report how much the annualized average moves, and state why a single figure is not enough.
  3. Pull a second venue’s public funding history, using a venue whose terms and location rules allow it for you. Confirm the unit and interval before comparing, and do not assume they match.
  4. Read openInterest and markPx from metaAndAssetCtxs each day for a week. Plot OI in coin terms and in dollars, and note any day on which the two disagree in direction.
  5. Write one paragraph on what your data do not show, including what they cannot say about who holds the positions.
Open Hyperliquid public API documentation →

Check yourself

1. Which statement correctly compares 0.00125% per hour and 0.01% per 8 hours?

0.00125% × 8 = 0.01%. A venue’s interval changes how a rate looks, and it must be normalized before comparison.

2. A 30-day funding sample sits at exactly the baseline for 70% of hours. What is the most careful reading?

The baseline is what the formula produces across a range of premiums. It does not measure balance or volatility.

3. A position is short perpetual and long spot at equal size, on one venue. The price rises sharply. What risk is most immediate?

Profit on spot does not automatically post as margin on the short unless the venue allows it. The short can be liquidated, leaving an unhedged position.

4. OI is unchanged after a trade. What happened?

When one party opens and the other closes, a position changes hands. Volume rises and OI does not.

5. Dollar OI falls by 10% during a day. Which further data point is needed before saying positions closed?

A dollar figure moves with price. Coin-based OI separates position changes from price changes.

6. Why can’t OI by itself tell whether longs or shorts dominate?

Aggregate long and short positions are equal by construction, as the CFTC’s notes state.

7. A carry position earns a 10.95% annual funding rate on notional. Which is the most accurate statement of its return?

The worked example showed 10.95% falling to about 6% on capital after costs, before accounting for any risk event.

8. Why might funding on two venues differ for a long time?

He et al. point to segmented margin and liquidity differences. Design differences and counterparty risk add to them.
← 13. Liquidation, insurance funds and auto-deleveraging 15. Reading public derivatives positions →

Describes past on-chain activity and how our data is built. Not investment advice. Method details: data methodology. Live panels: tracked whale wallets, updated 2026-11-05 15:41 UTC.