The Perpetual Futures Basis: Microstructure and Capture

Introduction

A perpetual future is a futures contract with no expiry. It stays close to the spot price because one side pays the other a funding rate at set intervals. When the contract trades above its spot index, longs pay shorts. When it trades below, shorts pay longs. Perpetuals trade far more than spot. He, Manela, Ross and von Wachter report median daily perpetual volume of $101.9bn in 2022, about two to three times spot volume on the same exchanges, using CoinGecko data.

The position that receives funding is short the perpetual and long the spot asset. It is often called market-neutral carry. That label hides most of what matters. Funding is a floating rate, set by each venue's own rules and known only after it has been paid. The position has no date on which its two legs must converge.

This paper studies funding on BTC and ETH perpetuals at Binance, OKX, Bybit, Deribit and Kraken, from November 2018 to 25 October 2026, using public exchange data. We measure its level, its persistence, its link to trend and volatility, its dispersion across venues and contracts, and its behaviour in nine stress episodes. We set these against the US cash rate, the visible short side of the market and the literature. We then describe, in general terms, a framework for holding the position and the risks it carries.

The paper reports market statistics only. It contains no strategy returns, simulated or otherwise. Funding levels are rates paid between market participants, before fees, margin costs, changes in the basis and venue risk. They are not the return of any position. Annualised figures are funding rates scaled to a year so that venues and periods can be compared.

Hedgehog Capital is an AI-first global macro research-and-risk service for professional and high-net-worth investors. Its services are not available to retail investors or to US citizens or residents. Clients keep their capital on their own exchange accounts, where each client's dedicated Hedgehog AI trades. Hedgehog never holds client assets and is not a fund. This paper describes Hedgehog's own approach only as far as Hedgehog has already published it.

How funding is set

Robert Shiller proposed perpetual futures in a 1992 Cowles Foundation paper for assets whose prices are hard to observe, such as real estate, with daily cash settlement that included an index of dividends or rents. Crypto reversed the purpose. Spot prices are visible every second, and funding pulls the contract towards them. He et al. date the first crypto perpetual to BitMEX in 2016. BitMEX ceased exchange operations on 23 September 2026. A contract with no expiry still ends when its venue does.

Ackerer, Hugonnier and Jermann derive no-arbitrage prices for linear, inverse and quanto perpetuals. In their framework the futures price is the risk-neutral expectation of spot at a random future time, and the funding rules set how strongly the contract is anchored. Kim and Park treat the funding rate as the design tool that ties the contract to its target. Funding is an engineered rate, and its rules shape what it measures.

Binance, OKX and Hyperliquid use one formula. Funding equals the average premium of the perpetual over its spot index, plus the gap between a fixed interest term and that premium, with the gap clamped at 0.05% either way. The interest term is 0.01% per 8 hours. OKX and Hyperliquid document the formula. Binance's is as described by He et al.

The clamp creates a mass point. Whenever the premium sits within 0.05% of the 0.01% interest term, funding prints the interest term exactly. Paid every 8 hours for a year, that is 10.95% annualised. A run of 0.0100% prints shows only that the premium stayed inside a band. It does not show how demand for leverage moved within it. Bybit's data carry the same signature (Table 2), although its formula is not confirmed here from Bybit's documentation.

Other venues use other rules. Kraken resets funding on its multi-collateral perpetuals, including PF_XBTUSD, every hour. It takes the average of the middle 30 of 60 one-minute premium observations and divides it by 8, with no interest term, no dampening and a cap of 0.50% per hour either way. Deribit publishes hourly funding through its public API. Its data show no fixed interest term, and 17.92% of hourly BTC prints are exactly zero, which points to a band in which no funding is charged. Deribit's formula is not confirmed here from its own documentation. Design shows in the level. On Deribit, funding has sat structurally below the clamp venues: over its full sample its median BTC day was 2.17% annualised, against 8.07% on Binance BTCUSDT (Tables 2 and 7). Kraken's 12 months of public history show no such gap.

Intervals and caps differ, and they change. When an OKX contract settles at its cap or floor, settlement steps from every 8 hours to 4, then 2, then 1. It reverts only when every settlement in the previous 12 hours has been within 0.20% either way. Binance publishes adjusted caps, floors and intervals through its funding-info endpoint, and Bybit sets an interval per symbol. Table 1 summarises the rules.

Contract design adds a layer. A USDT-margined contract is priced and margined in a stablecoin. An inverse contract, such as Deribit's BTC-PERPETUAL, is quoted in dollars but margined and settled in bitcoin. Exchanges do not cross-margin with one another, so spot on one venue and a perpetual on another must each be funded and margined separately (He et al.). The stress episodes below show that these details decide outcomes.

Background and evidence

The literature explains the basis as leveraged demand meeting scarce arbitrage capital. He et al. study hourly Binance prices for five coins from January 2020 to March 2024. Deviations from no-arbitrage prices are much larger than in currency markets and move together across coins. They fall by about 22 percentage points a year, which the authors link to growing arbitrage capital. They rise after strong spot returns, which they read as extrapolative demand for leverage. They compressed after Binance introduced portfolio margining on 20 April 2022. Citing news reports, the authors add that Alameda Research and Three Arrows Capital appear to have moved from this arbitrage into directional bets around late 2021 and early 2022.

He et al. also stress that the position carries risk before costs. A perpetual has no set expiry, so nothing forces the legs to converge when an arbitrageur needs to unwind. The interest term is a convention rather than a market rate. The stablecoin interest rate they use as the arbitrageur's cash rate, an average of Aave rates on USDT, USDC and DAI, was below it in 93% of their sample hours. The median daily effective spread of the BTC perpetual was 0.11 basis points, but on 12 March 2020 it reached 5.95.

A BIS working paper by Schmeling, Schrimpf and Todorov reaches a similar view. Carry, the annualised gap between futures and spot, averages above 10% a year and can reach about 60%. Interest rate differentials explain very little of it. Most of the variation comes from a volatile convenience yield, driven by smaller investors chasing trends and by a shortage of capital willing to take the other side. The arbitrage is risky because margins and liquidations spike in market falls. High carry predicts future crashes and tends to rise with the price of crash insurance.

Christin, Routledge, Soska and Zetlin-Jones use Binance minute data from August 2020 to June 2022 and find that derivative volume is driven by the long side. They link the carry to the price long traders pay for leverage, and point to Binance's cut in maximum leverage from 125x to 50x in July 2021. Makarov and Schoar document large, recurring price gaps across crypto exchanges, wider across countries than within them and wider when bitcoin rises quickly.

The dated-futures basis tells the same story on a regulated venue. After the SEC approved the listing and trading of US spot bitcoin exchange-traded products on 10 January 2024, the long leg could be held in a listed product. The Banque de France reports institutional holdings of bitcoin ETFs above $33bn at end-2024, nearly $18bn of it held by hedge funds, which held net short futures positions as part of this arbitrage. It puts the one-month basis at 0.3% to 0.8%, or 5% to 12% annualised. CF Benchmarks reports that the annualised CME front-month basis approached 25% in February 2024, briefly dipped below zero in March 2025 and was near 10% by May 2025. It finds the basis driven mainly by price momentum and sentiment.

The closest traditional analogue is the Treasury cash-futures basis. Barth and Kahn find that it grew after 2016 to as much as half of hedge funds' Treasury positions. In mid-March 2020 volatility and margins jumped. The BIS describes forced selling by leveraged investors exploiting small yield differences. The New York Fed's Lorie Logan said margin changes likely contributed to a partial unwind, and noted that the Fed bought about $1.6 trillion of Treasuries between 13 March and 10 June 2020. A Federal Reserve note describes the trade as exposed to both futures margins and repo spreads. A Dallas Fed article, citing Barth and Kahn (2025), links roughly 60% of the rise in hedge funds' Treasury positions over recent years to it.

Crypto has the same shape with a different backstop. No central bank stands behind crypto venues in a crash. The backstops are each venue's insurance fund and, behind it, auto-deleveraging, which closes positions on the winning side.

Data and method

We downloaded funding histories from the public APIs of Binance (USDT-margined from 2019, coin-margined from 2020), OKX (from 31 August 2021), Bybit (inverse from 2018, USDT-margined from 2020), Deribit (hourly, from 30 April 2019) and Kraken (hourly, from 15 October 2025, the limit of its public history). Spot prices are Coinbase Exchange daily closes. Implied volatility is Deribit's DVOL index. The cash rate is the 3-month Treasury bill from FRED, quoted on a discount basis. CME positioning is from the CFTC Traders in Financial Futures report. USDe supply is from DefiLlama. Data run to 25 October 2026 (UTC).

Each print is annualised by multiplying it by 8,760 and dividing by the hours it covers, so a 0.0100% 8-hour print is 10.95%. Annualisation is simple, not compounded. Daily funding is the funding paid in periods ending that UTC day, divided by the hours those periods cover and annualised. The composite is the equal-weighted mean of daily funding on the USDT-margined contracts at Binance, OKX and Bybit and the inverse contract at Deribit, with at least two venues required. Its membership grows over time. It holds Binance and Deribit only in 2019, adds Bybit from March 2020 for BTC and October 2020 for ETH, and adds OKX from 31 August 2021. OKX has no data from 30 September 2021 to 16 January 2022. A fixed three-venue BTC panel of Binance USDT-margined, Bybit inverse and Deribit funding shows the same compression, from 26.32% in 2021 to 2.84% in 2026. Persistence is summarised by a first-order autoregression with an intercept. The half-life is ln(0.5) divided by ln(phi), where phi is the autoregressive coefficient. Weekly and monthly half-lives are converted to days. Realised volatility is the standard deviation of daily log returns over the previous 30 days, annualised with the square root of 365. Regressions use Newey-West standard errors with 30 lags.

Four caveats apply throughout. First, venues changed their rules during the sample, including caps, intervals and, at OKX, the formula itself. OKX's notice of 25 January 2024 gave its funding rate as a capped average premium with the interest rate set to zero (OKX). Its archive shows no print at exactly 0.0100% before 31 May 2024 for ETH and 24 April 2025 for BTC. The date of the later move to the clamp formula is not confirmed from OKX's documentation. Second, Deribit and Kraken have no interest term, so venue comparisons mix design and market effects. Third, the panel includes only surviving venues, those that still publish data in October 2026. FTX, which failed in November 2022, and BitMEX, which closed in September 2026, are absent. The early composites therefore rest on two or three venues and may understate dispersion and stress. Fourth, full-sample averages are dominated by 2020 and 2021. Every statistic is computed by a script from saved raw files and can be reproduced.

Findings: level, design and the clamp

Table 2 shows the full-sample distribution of daily funding. Three features stand out. Means sit well above medians, because a few high-funding episodes carry the average. Binance BTCUSDT averaged 11.46% annualised with a median of 8.07%, and its 99th-percentile day was 110.23%. The clamp venues carry a large mass at exactly 0.0100% per 8 hours, which held for 34.85% of Binance BTCUSDT prints and 40.27% of Bybit BTCUSDT prints. And Deribit sits lower: its median BTC day was 2.17% annualised.

Table 3 tracks the cross-venue composite by year. BTC fell from 28.36% annualised in 2021 to 2.83% in 2026 to 25 October, and ETH from 32.42% to 2.19%. The share of negative days rose in 2026 to 19.5% for BTC and 24.2% for ETH, after three years at 6.0% or less for BTC.

Table 4 shows how the clamp and the upside tail have changed on Binance BTCUSDT. In 2019, 66.7% of prints sat exactly at 0.0100%. That share fell to 6.6% in 2026, while the share of negative prints rose to 25.5%. Prints above the interest term have stopped. No 8-hour print on Binance's BTCUSDT or ETHUSDT contract has exceeded 0.0100% since 16:00 UTC on 9 December 2024, a run of 2,055 consecutive prints. Bybit BTCUSDT and OKX BTC-USDT-SWAP last printed above it at 00:00 UTC on 23 July 2025. On each of these series the highest print since has been exactly 0.0100%. In 2021, 50.7% of Binance BTCUSDT days averaged above 10.95% annualised. In 2025 and 2026 none did.

On Deribit the tail has thinned but not vanished. Deribit BTC funding averaged above 10.95% on 41.1% of days in 2021, 35.5% in 2024, 17.5% in 2025 and 2.7% in 2026. The data establish that high prints have been absent on the clamp venues. They do not establish why. Abundant short supply, changes to how venues build their premium index, or both could explain it.

A headline average funding rate therefore mixes market demand with venue design. On a clamp venue, a premium near zero prints 10.95% annualised by construction. On Deribit, where 17.92% of hourly BTC prints are exactly zero, the same premium prints at or near zero. Any comparison of funding across venues, or with the cash rate, has to start there.

Findings: funding against the cash rate

Table 5 sets the 30-day average of the composite beside the 3-month Treasury bill rate. It compares two published rates. It is not a return. It leaves out fees, margin, changes in the basis, and the fact that stablecoin collateral does not generally earn the bill rate.

In 2021 the BTC composite's 30-day average was 29.51% against a bill rate of 0.05%. In 2026 to 25 October it was 2.84% against 3.70%, and it sat below the bill rate on 55.4% of days. For ETH the share was 77.9%. The bill rate was 4.04% on 22 October 2026, the latest FRED observation. FRED quotes it on a discount basis, slightly below the bill's investment yield.

Findings: persistence

Funding mean-reverts quickly from day to day but slowly at the level of regimes. Daily funding on the BTC composite has a half-life of 3.5 days, and single BTC contracts in Table 6 range from 1.8 days (Deribit) to 4.1 days (Binance). Weekly and monthly averages are far more persistent. For the BTC composite their half-lives are 13.1 and 39.3 days. Across the venues in Table 6 they range from 8.6 to 14.7 days and from 18.2 to 59.3 days. The BTC composite's autocorrelation is still 0.196 at 90 days and 0.052 at 180 days.

The regime is persistent and the daily print is noisy around it. A single print says little. A weekly or monthly average says much more about the following weeks. Deribit has the shortest half-lives for each asset at all three horizons, probably in part because the clamp holds the other venues' prints at the interest term in quiet markets. These are linear summaries of a series with a mass point and shifting regimes. They describe typical behaviour and are not a model to trade on. The monthly estimates rest on 60 to 91 monthly means, and the uncertainty is wide: the 95% interval for the BTC composite's monthly half-life runs from roughly 24 to 77 days.

Negative regimes can last. Binance BTCUSDT's 30-day average was negative for 84 consecutive days, from 27 February to 21 May 2026, the longest such stretch in its seven-year history, with a low of -3.55% annualised on 9 May. Binance ETHUSDT's was negative for 96 days, from 12 February to 18 May 2026, and Deribit ETH's for 189 days to 19 December 2022. The BTC composite's 30-day average was not negative on any day from 2023 to 2025; in 2026 it was negative on 11.1% of days.

Findings: venues and contracts

Table 7 compares BTC funding across the four primary venues from 16 January 2022, when OKX's data resume after a gap that began on 30 September 2021. Binance and Bybit averaged 6.54% and 6.41% annualised. Deribit averaged 4.60% with a median of 1.53%, and it was the lowest venue on 41.7% of days, which reflects its formula as much as its market. No venue led most of the time.

Venues share the funding regime more than the daily move. Daily BTC funding levels correlate between 0.566 (Bybit and Deribit) and 0.820 (Binance and OKX). Daily changes correlate only 0.227 to 0.529, and all four venues share a sign on only 59.7% of days. The mean daily range across them is 9.76 percentage points, with a 95th percentile of 21.62 and a maximum of 180.9 on 12 March 2023. ETH is more tightly linked, with a mean pairwise correlation of 0.768 in levels.

The gap between venues has narrowed. The mean BTC range fell from 12.58 percentage points in 2022 to 5.30 in 2026, and its 95th percentile from 28.93 to 9.83. A longer three-venue panel, using Binance USDT-margined, Bybit inverse and Deribit, shows a mean range of 31.02 percentage points in 2021 and 4.98 in 2026. Narrower dispersion means the venue matters less to the funding rate observed, and part of what remains is formula rather than mispricing. The range is measured after the fact, across venues, before fees, margin and transfer costs.

Kraken's public history covers only the 376 days from 15 October 2025. Over that period its BTC funding averaged 3.43% annualised, against 3.34% on Binance, 3.46% on OKX, 2.74% on Bybit and 2.34% on Deribit, with correlations of 0.484 to 0.612 to the other venues. Kraken ETH averaged 3.57%, above the 1.60% to 2.55% on the other four. Kraken has no interest term, so its history shows that a venue without one need not sit below the clamp venues.

Contract type matters more than the averages suggest. On the same venue, USDT-margined BTC funding exceeded coin-margined funding by 1.66 percentage points on average at Binance and 1.84 at Bybit, but the median gap on both was zero. The difference comes from a few episodes, which the stress section identifies.

Findings: trend, volatility and what followed

Funding follows the price trend (Table 8). For BTC the 7-day composite correlates 0.528 with the trailing 30-day return and 0.026 with trailing 30-day realised volatility. For ETH the figures are 0.564 and 0.052. A regression with Newey-West standard errors finds the trailing return significant for both assets and volatility insignificant, with a downward trend of 1.84 percentage points a year for BTC and 2.00 for ETH. In the lowest fifth of BTC 30-day returns the 7-day composite averaged 0.79% annualised. In the highest fifth it averaged 26.45%. The weak volatility correlation hides a two-sided link. Both the lowest and the highest funding have come in volatile markets. Trailing BTC volatility averaged 69.6% on days in the lowest fifth of funding and 63.5% in the top fifth, against 46.0% to 50.2% in the middle fifths. Sorted the other way, the 7-day composite averaged 5.36% in the calmest fifth of BTC days, against 14.73% and 12.76% in the two most volatile fifths.

This matches the literature. He et al. find that deviations rise after high past returns, the BIS paper attributes carry to trend-chasing, and CF Benchmarks finds the CME basis tracking momentum. Funding is the price of leveraged long exposure, and demand for it rises in rallies. Funding is only loosely tied to implied volatility: the 7-day BTC composite correlates 0.195 with Deribit's DVOL index. DVOL itself exceeded trailing 30-day realised volatility on 77.3% of BTC days from March 2021, by a mean of 7.7 percentage points. That gap is the subject of Volatility Risk Premia in Crypto Options, a Hedgehog paper in preparation.

High funding has come before higher volatility (Table 9). When the 7-day BTC composite was in its top fifth, above 13.1% annualised, realised volatility over the next 30 days averaged 74.0%, against 50.0% to 53.3% in the other groups. For ETH the figures were 100.2% against 65.2% to 73.6%. The pattern comes mainly from the early years. Days from 2019 to 2021 make up 365 of the 514 BTC top-fifth days. From 2022 to 2026, BTC top-fifth days were followed by volatility of 55.1%, against 43.9% to 49.1% for other days. For ETH in that period there was no such gap: 64.0% against 57.2% to 70.9%.

These are descriptive statistics on overlapping windows from a single history, and the fifths use cut-offs that were not known at the time. They are not a trading rule and should not be read as one. The BIS paper finds that high carry predicts future crashes. The relevance here is to risk. The conditions in which funding has been highest have also been those in which a hedged position was most likely to be tested.

Findings: stress episodes

Table 10 covers nine stress windows chosen by us. The main lesson is that the severity of funding stress depends on who is positioned where, and in which contract, more than on the size of the price fall.

In March 2020 BTC fell 52.8% peak to trough and funding hit venue floors. Binance BTCUSDT printed -0.300% for the 8 hours to 08:00 UTC on 13 March, and Bybit's inverse contract -0.375%. Recovery was slow. Binance BTCUSDT's 7-day average took 58 days to regain half its pre-event 30-day average; Bybit's inverse contract and Deribit took 138 and 136 days. In May 2021 Binance's USDT-margined BTC contract bent but held, with its lowest day at -16.0% (Bybit's reached -50.1%), while Binance's coin-margined BTC contract printed -0.300%. Pre-event 30-day funding of 24.74% to 62.28% annualised across BTC series gave way to averages of -5.37% to 5.68% over the following 30 days, and the regime changed. The IMF judged that the May falls were likely exacerbated by leverage and automatic liquidations, and noted exchange outages on 19 May, when liquidations peaked.

FTX was a venue failure as well as a price fall. BTC fell 26.6%, but funding turned deeply negative on every venue in our panel for two to six days. It was below -20% annualised for two days at Binance and OKX, four at Bybit and six at Deribit, where funding stayed negative from 8 to 16 November. Binance BTCUSDT printed -0.1192% for the 8 hours to 00:00 UTC on 10 November 2022, and Deribit averaged -129.4% annualised on 11 November. He et al. read the pattern as FTX users closing short futures while perpetuals on FTX traded above spot. The CFTC alleged that more than $8bn of customer deposits were lost and that undisclosed code features benefited Alameda Research. In the bankruptcy, the court held that digital-asset claims must be valued as of the 11 November 2022 petition date, and accepted the Debtors' method of converting them to US dollars. In effect, a short perpetual on the failed venue became a fixed dollar claim, while a spot leg held elsewhere stayed exposed to the price.

The USDC de-peg of 10 to 13 March 2023 isolated collateral risk. BTC fell only 9.9% peak to trough, to a low at 11:00 UTC on 10 March, then rose 32.5% to about 26,000 by 16:00 UTC on 14 March. Binance BTCUSDT's lowest day was -1.6%, and Bybit BTCUSDT's daily funding stayed positive. Coin-margined and Deribit funding plunged: Deribit BTC averaged -170.2% annualised and Bybit's inverse contract -114.1% on 12 March, and OKX's inverse contract -109.5% on 11 March. The cross-venue BTC range reached 180.9 percentage points, the widest of the 2022 to 2026 panel. A hedged position's funding depended on which contract and which collateral it used.

On USDT-margined contracts, the deepest daily ETH funding in our data came around the Merge in September 2022, not in a crash. Binance ETHUSDT printed -0.3019% for the 8 hours to 00:00 UTC on 15 September and averaged -239.7% annualised on 14 September, while BTC funding barely moved. Deribit's deepest ETH day remains 13 March 2020 (-545.2%). Market commentary attributes the Merge episode to holders shorting perpetuals ahead of the upgrade. We found no primary source for that account. In August 2024 BTC fell 23.7% around the yen carry unwind, which the BIS describes as volatility amplified by procyclical deleveraging and margin increases, yet Binance BTCUSDT's lowest day was only -0.4%.

On 10 October 2025 BTC fell 6.6% on Deribit's hourly index, which misses intraday lows, and by 21:00 UTC on 11 October it was 9.4% below its 10 October high. The 14.5% in Table 10 runs from the 6 October high to 16 October. FTI Consulting estimates that more than $19bn of crypto leverage was liquidated in about a day and that top-of-book BTC depth fell by more than 90% on key venues. USDT-margined BTC funding moved little. Binance BTCUSDT's lowest print in the window was -0.0109%, and the lowest day on OKX's USDT-margined contract was -13.0%. Deribit BTC's hourly funding turned sharply positive, reaching 346.0% annualised in the hour to 23:00 UTC as its index fell from 121,602 at 12:00 to 113,780 at 22:00. ETH funding fell sharply. For the 8 hours to 00:00 UTC on 11 October, OKX ETH-USDT-SWAP printed -0.1461%, Bybit ETHUSDT -0.1177% and OKX's inverse ETH swap -0.2896%.

For hedged positions the main threat that day came from auto-deleveraging rather than funding. Chitra reports that Hyperliquid used it to close $2.1bn of positions in 12 minutes. His method suggests that it took $45.0m to $51.7m more from profitable traders than the shortfall required, and that Binance overused it far more than Hyperliquid. FTI Consulting reports that some of the best-hedged shorts were reduced or closed, which turns a hedged book into an unhedged one, and that frozen interfaces stopped some traders from managing risk.

Findings: the short side of the market

Part of the short side is visible. The CFTC's explanatory notes describe Leveraged Funds as typically hedge funds and various types of money managers, whose strategies may involve outright positions or arbitrage. Their net position in standard CME bitcoin futures (Table 11) went from -9,831 contracts at end-2023 to -23,197 contracts, about $12.31bn, on 17 December 2024. That was the largest net short in the series since April 2018, and open interest also peaked that week. It was a one-week peak. Open interest rose from 37,588 contracts on 10 December to 42,524 and fell back to 37,125 on 24 December, and the net short was -20,986 and -21,102 contracts in the weeks either side. By 20 October 2026 it was -4,942 contracts, about $2.05bn. In CME ether futures the net short peaked at about $4.46bn on 7 October 2025 and stood at about $0.63bn on 20 October 2026. These net shorts are compatible with basis trading against spot ETFs, as the Banque de France describes, but they do not prove it.

Ethena's USDe offers a public view of the perpetual trade held at scale. Its documentation describes backing that pairs volatile assets with short derivatives, historically concentrated in spot crypto hedged with short perpetuals, and says the mix is now diversifying. USDe supply rose from $0.09bn on 1 January 2024 to a peak of $14.82bn on 4 October 2025. It fell 14.1% between 10 and 13 October 2025, and on 25 October 2026 it was $4.59bn, 69.0% below the peak. The New York Fed explains the mechanism. When funding turns negative, holders redeem, and the protocol must buy back its short futures and sell ether, pressing on both markets at once. It reports that USDe's market value fell by more than 13% after the October 2025 funding reversal, and that USDe briefly traded at $0.65 on Binance while its peg held elsewhere.

The short side has grown and shrunk with the carry, and it can be forced out at the moment funding reverses.

A systematic framework for holding the position

The findings point to a framework in five parts. This section describes it in general terms. It is a way of managing a position that receives or pays a floating rate and whose risks sit in the tails. It does not describe Hedgehog's implementation. It is not a claim that the position will be profitable, and it contains no estimate of what it would have earned.

What is held. The position is long the spot asset and short the perpetual on the same underlying, sized so that price moves largely offset. It receives funding when funding is positive and pays it when negative, and it is exposed to the gap between perpetual and spot, to the venues holding each leg and to the collateral posted.

1. Measure. Normalise prints for interval, formula and contract before comparing them. Separate the clamp's mass point from genuine premium. Track a cross-venue composite, the dispersion across venues, and funding averaged over weeks and months as well as single prints.
2. Condition. The evidence supports conditioning on four things: the level and persistence of funding at regime frequency; the price trend, because funding rises in rallies; the spread over the cash rate; and signs of stress, because in 2020 and 2021 the highest-funding conditions came before higher volatility, and stress turns funding negative and drives venues apart.
3. Choose the venue and contract. The highest-paying venue is not the best by default. Formula, fees, margin rules, collateral currency, index construction, auto-deleveraging rules and venue credit all differ, and in March 2023 a USDT-margined and an inverse contract on the same venue behaved very differently.
4. Size within limits. Size to the regime rather than the latest print, with limits that tighten as measured risk rises. Keep the short leg's liquidation price well away from the market, because a spot leg on another venue cannot meet its margin calls. Limit exposure to any single venue and any single stablecoin.
5. Plan for stress. Decide in advance how the position is reduced when funding stays negative for weeks, when venues diverge, when auto-deleveraging closes the hedge or when a venue fails. Interfaces can freeze, so the plan must not depend on acting at the worst moment.

Persistence is the most useful single property. Because the daily print is noisy and the regime persistent, a framework that responds to weekly or monthly averages reacts to changes of regime rather than to noise. The cost is lag: a regime can turn faster than a 30-day average shows, as it did in March 2020 and in November 2022. Conditioning therefore needs a fast stress input beside the slow regime input. The trend link adds a warning. Funding is highest when prices have risen most, and in 2020 and 2021 those days were followed by higher volatility. Sizing up with funding, without a check on stress, concentrates risk at the wrong time.

Venue choice needs the same care. Deribit's documentation shows that its inverse perpetuals, such as BTC-PERPETUAL, are outside both its Liquidity Support Program and its auto-deleveraging. Hyperliquid closes profitable opposing positions at the previous mark price, ranked by profit and leverage. These differences belong in the venue decision alongside the funding level.

How Hedgehog's published framework fits. Hedgehog's strategy library lists perpetual futures funding rate capture, spot-futures basis across multiple venues and cross-asset carry, with regime-conditioned sizing that is larger when funding is persistent. Its published note on basis strategies describes conditioning the trade every day on three inputs: the current funding rate across exchanges, the persistence of that rate, and a cross-exchange basis comparison to identify the best execution venue. When funding is high and persistent the position is sized up; when funding compresses or reverses it is reduced.

That sizing sits inside Hedgehog's risk limits. Size has two limits, and the smaller applies: the client's allocation and the most that strategy can take. EVAR, Hedgehog's loss estimate conditioned on the current market regime, is calculated for each strategy and for the book. It updates in real time as prices and news move, and a strategy's position limit tightens when its EVAR rises. Research and the regime estimate are refreshed every day. Clients keep their capital on their own exchange accounts at Kraken, Coinbase, Deribit, Binance or Hyperliquid. Each client's dedicated Hedgehog AI trades there through a read-and-trade key with no withdrawal permission, and the client sets which strategies run and at what size. Hedgehog is preparing papers on the regime and portfolio frameworks: Regime Detection in Digital Asset Markets and Multi-Strategy Portfolio Construction Under EVAR Constraints. This paper does not describe the parameters, data or models behind any of these steps.

Implications for allocators

Five implications follow for professional allocators.

1. Funding needs a full specification. A quoted funding level needs a venue, a contract, a collateral currency, a formula and a period. Full-sample averages are carried by 2020 and 2021, and on clamp venues medians are pulled towards 10.95% by construction.
2. Funding is a floating rate, not a yield. The BTC composite averaged 28.36% annualised in 2021 and 2.83% in 2026 to 25 October, and its 30-day average has been below the Treasury bill rate on most days of 2026.
3. The regime says more than the print. Monthly averages have half-lives of several weeks; daily prints have half-lives of days. Binance ETHUSDT's 30-day average stayed negative for 96 days in 2026.
4. The tails need a plan. Reversal, auto-deleveraging, venue failure and collateral stress all appeared in the sample. Any implementation should be able to show what happens to the position in each.
5. Custody and venue risk are separate questions. Keeping capital on one's own exchange accounts avoids pooling it with other investors, but it does not remove exchange counterparty risk. FTX showed that a short leg on a failed venue becomes a claim while the spot leg stays exposed.

Questions that test any implementation of the position include the following.

• The venues and contracts used for each leg, and the collateral posted in each.
• How the short leg is margined, and how far it sits from liquidation in a fast rally.
• Which venues apply auto-deleveraging to the contracts used, and what happens to the position if they do.
• How sizing responds to a negative funding regime that lasts months.
• The limits per venue and per stablecoin, and how they tighten as measured risk rises.
• How funding is compared with the cash rate after fees, margin and transfer costs.

Risks and limitations

The position carries risks that the funding level does not show.

• Reversal. Funding can turn negative and stay there, as it did for 84 days on Binance BTCUSDT's 30-day average in 2026.
• Crowding. The short side grows and shrinks with the carry, and redemptions or margin calls can force synchronised unwinds.
• Liquidation of the short leg. In a rally the short perpetual loses margin while the spot leg gains elsewhere, and venues do not cross-margin.
• Auto-deleveraging. In a crash a venue may close the short leg and leave the spot leg unhedged, as on 10 October 2025.
• Venue failure. A failed venue freezes the leg it holds, as FTX did in November 2022.
• Collateral, index and oracle. Stablecoin and coin margin each carry their own risk, as March 2023 showed, and a venue that marks against its own thin market can print prices that diverge from elsewhere.
• Operations and rules. Exchanges reported outages when liquidations peaked in May 2021. Caps, intervals and formulas change, venues close, and no lender of last resort stands behind them.
• Models and connections. Strategies, risk limits and AI models can fail. Markets can move faster than risk limits can respond, and API connections, hardware and networks can fail.

The analysis has limits of its own. Exchange data are as published by each venue, without independent audit. Rules changed during the sample, including OKX's formula. The funding formulas of Binance, Bybit and Deribit are not confirmed here from the venues' own documentation. The panel includes only surviving venues: FTX and BitMEX are absent, OKX data before September 2021 were not available, and Kraken's history covers about 12 months. The early composites therefore rest on two or three venues and may understate dispersion and stress. Daily data hide intraday extremes, and the hourly Deribit index misses the lows of 10 October 2025. The stress windows are our choice. The statistics describe association and do not establish cause. Table 9 rests on overlapping windows from one history, with fifths set on the full sample, and its pattern weakens after 2021. CFTC categories describe types of trader and do not show their trades. DefiLlama is not an official statistic. None of the figures is the result of a strategy, and none should be converted into one.

Conclusion

The perpetual futures basis is the market price of leveraged long exposure, paid through rules that each venue writes. Those rules explain a large part of what headline funding shows. On clamp venues a premium near zero prints 10.95% annualised; on Deribit it prints at or near zero. Funding rises with the price trend, persists over weeks and months, and has compressed since 2021. In 2026 its 30-day average has sat below the US cash rate on most days, and no print on Binance's main contracts has exceeded the interest term since December 2024.

The averages hide the main risks. Funding can stay negative for months. Venues and contracts diverge most in stress. Auto-deleveraging and venue failure can break the hedge itself. A systematic framework is therefore less about finding the highest rate than about measuring funding correctly, conditioning on regime and stress, choosing venue and contract with care, and sizing within limits that tighten as risk rises. Any implementation is best judged on how it handles the tails.

About the data

Data through 25 October 2026 unless stated. Funding histories were downloaded on 26 October 2026 from the public APIs of Binance (USDT-margined BTCUSDT and ETHUSDT from September and November 2019; coin-margined BTCUSD and ETHUSD from August 2020), OKX (USDT- and coin-margined BTC and ETH swaps from 31 August 2021, from its monthly archives and funding-rate history endpoint, with no prints from 30 September 2021 to 16 January 2022), Bybit (USDT-margined BTC and ETH from March and October 2020; inverse BTC and ETH from November 2018 and January 2019), Deribit (hourly BTC-PERPETUAL and ETH-PERPETUAL funding and index prices from 30 April 2019) and Kraken (hourly PF_XBTUSD and PF_ETHUSD from 15 October 2025, the limit of its public history). Spot prices are Coinbase Exchange daily closes; implied volatility is Deribit's DVOL index from 24 March 2021; the cash rate is FRED series DTB3 (discount basis), forward-filled over weekends; CME positioning is the CFTC Traders in Financial Futures report, futures only, for the standard bitcoin and ether contracts, to 20 October 2026, with US dollar values at the Coinbase close on each report date; USDe supply is from the DefiLlama stablecoins API; venue caps, floors and intervals are a snapshot at 14:49 UTC on 26 October 2026. Data run to 25 October 2026 (UTC); for 8-hour venues the last day, 25 October, covers prints to 16:00 UTC. Each print is annualised (simple, not compounded) as rate times 8,760 divided by the hours it covers; daily funding is the funding paid in periods ending each UTC day, divided by the hours those periods cover and annualised; the composite is the equal-weighted mean of Binance, OKX and Bybit USDT-margined and Deribit inverse funding, with at least two venues required, and its membership grows from two venues in 2019 to four continuously from 16 January 2022; persistence uses OLS first-order autoregressions with an intercept on daily, weekly and monthly means; realised volatility is the trailing 30-day standard deviation of daily log returns annualised with the square root of 365; regressions use Newey-West standard errors with 30 lags; Table 9 fifths are set on the full sample. The panel includes only venues that still publish data in October 2026: FTX and BitMEX data, OKX data before September 2021, Kraken data before October 2025 and CME futures prices were not available. Every statistic is computed by scripts from saved raw files and is reproducible. All figures are market statistics or are cited from the sources listed; no strategy return, simulation or backtest has been computed.