Every Long Has a Short: Reading Open Interest, Funding and Basis

Every derivatives contract that's open right now has two owners. Somebody is long it and somebody is short it, and neither side can exist without the other. It sounds too obvious to write down, and yet almost every misreading of derivatives data I've watched people make, me included, starts from forgetting it. "Open interest is rising, so the longs are piling in." "Funding is positive, so everyone's bullish." "Liquidations hit a record, so we know how much leverage got flushed." Each of those sentences treats one side of a pair as the whole of it, or treats a number as a census when it's a price, a sample or a model.

This post is about reading those numbers. It's a preamble, like Turn the Candles Off was for volume, and it exists for the same reason: the derivatives panel coming as Tool 02 in Applied Market Mechanics reads open interest, funding, basis and order flow, and rather than explaining each of them inside the Tool, they get one page here that the Tools can point back to. It sits between two posts that already exist. Part 5 of Understanding Market Mechanics explains the machinery: the curve and the basis, funding and its formula, the mark price, the liquidation engine. Read the Spec Sheet explains the contracts: inverse against linear, and what one contract is actually worth. Neither of them says what the numbers those contracts publish are counting, and that's this page.

Every figure here follows ether from the first full week of CME's ether futures in February 2021 to 1 December 2023, and every one of them carries the same three dashed lines: 19 May 2021, 13 June 2022 and 9 November 2022. They're the lines I drew on a chart I posted in reply to a question about the CME premium on 1 December 2023, on the three days volume spiked hardest: the May 2021 crash, the Celsius and Three Arrows capitulation, and FTX. Seeing the same three days through every number is the point. Each number counts something different, so each one tells the story of those days differently, and that chart of mine turns out to be one of the worked examples, for a reason I'll get to.

#Open Interest

Open interest is the number of contracts open at a moment: positions that have been opened and not yet closed, offset or settled. A trade between two people who are both opening a position adds one contract to it. A trade between two people who are both closing removes one. A trade where one side opens and the other closes, a new buyer taking over from a seller who's leaving, changes who holds the contract and leaves the count where it was. I'll write it as , in whatever unit the venue reports.

What it measures is how much is at stake, in the same way volume measures how much business got done. What it can't measure is direction, because every contract in the count is a long and a short at once. There's no such thing as open interest that's mostly longs. When someone says rising open interest during a rally means the longs are piling in, they're describing a market where every one of those new longs has a new short on the other side of it, and the claim is really about which of the two was more eager to get in, which is a different number (aggressor flow, further down).

You'd think the counting itself would at least be settled. It wasn't, at one of the larger venues, until this year.

Bybit's API returns open interest in two fields. The original one, openInterest, counted the long and the short of each contract separately, and on 11 June 2026 Bybit switched its displays to single counting and added a singleOpenInterest field alongside. Pull both for its ether perpetual across this window and the first is exactly twice the second, on every one of the 1,036 days from February 2021 to early December 2023. The doubling was all in the counting. Anyone who compared Bybit's open interest with another venue's over those years was comparing a pair count with a contract count, and anyone who added venues together to get "total open interest" was adding the two.

So the first check on any open interest figure is what it counts: contracts, or sides of contracts.

#Who's on Each Side

Exchange open interest tells you how many pairs exist and nothing about who's in them. For regulated US futures there's one public source that does, and it shows the pairing more plainly than anything an exchange publishes. Every week the CFTC's Traders in Financial Futures report breaks a market's open interest into four kinds of holder (dealers, asset managers, leveraged funds and other reportable traders, with small traders as the remainder) and lists each one's longs and shorts. Its explanatory notes describe who falls where. Add every category's longs together, then every category's shorts, and you get the same number twice, and that number is the open interest, every week without exception. The report is the "every long has a short" rule written down as an accounting identity.

For CME ether, the leveraged funds were net short in 124 of the 139 weeks here, and by the last week of November 2023 they held about 180,700 ether short against 100,800 long, while asset managers held 79,300 long against 2,000 short. That fits the cash-and-carry trade Part 5 describes, where a fund buys spot and sells the future to collect the basis, and it's a reason not to read leveraged funds' shorts as a bet on lower prices. The report shows positions. The reasons behind them are inference, and the same footprint can come from a hedge or a view.

Two things to know before using it. The figures are contracts, and CME's ether future is 50 ether while its micro ether future is a tenth of one, so the two markets' counts can't be added until each is multiplied by its size, which is what the figure does. And the report is a Tuesday snapshot published on Friday, so it's a weekly reading of positions, three days stale on arrival.

#The Unit Moves Under You

The second check is the unit, and it catches people the same way it did in the volume post: a count of contracts, of coins or of dollars can each move for a reason that has nothing to do with positions being opened or closed.

A linear contract, margined in dollars, usually reports open interest in coins, and anything quoting it in dollars has multiplied by price. An inverse contract, margined in the coin, usually reports in dollars, because each contract is a fixed number of dollars. Convert either one to the other unit and you've multiplied or divided by the price, and the price is moving:

with the spot price.

Bybit's inverse ether perpetual had 568 million dollars of open interest on 10 November 2021, the day of ether's all-time high, and 147 million on 13 June 2022, a fall of 74%. Counted in ether, the same positions went from 122,600 to 121,300. Read in dollars, three quarters of the market left. Read in ether, almost nobody did, and what fell was the price of the thing being counted. Bybit's linear contract tells the mirror story over the same months: in ether its open interest more than doubled, from 37,700 to 84,200, while in dollars it fell from 175 million to 102 million.

Neither unit is wrong. They answer different questions. Ether tells you whether positions were opened or closed; dollars tell you how much money is exposed, which is what a liquidation engine cares about. The mistake is reading a change in one as if it were a change in the other.

#Long/Short Ratios

Most venues also publish a "long/short ratio", and since every long has a short, it's worth asking what it can possibly be counting. The answer is accounts. Binance's version, for example, compares the number of accounts holding net long positions with the number holding net short, and a variant restricted to its largest traders compares their positions. A ratio of two to one means twice as many accounts are long as short, which is entirely compatible with the shorts being fewer, larger and better funded. The contracts still balance one for one. Read these ratios as a headcount of a crowd on one venue. As a measure of which way money leans they say nothing, because the money in every contract is matched one for one.

#Funding

Part 5 covers what funding is and how it's built: a payment between the two sides of a perpetual, every interval, from whichever side the perpetual's price is leaning towards to the other, set by the formula

with the premium index and the fixed interest rate of per eight hours. Three things about reading it follow from that formula, and a fourth from the clock.

The first is that funding is a price, paid by one side to the other, and it says where the perpetual traded relative to its index, averaged over the interval. It doesn't say how many people are long or short, for the reason the whole of this post keeps coming back to.

The second is that it's blind in the middle. Because of the clamp, any premium between and prints exactly . A market where the perpetual sat perfectly on its index and a market where it sat noticeably above it produce the same number, and you can't tell them apart from the print.

Across this window, 36.7% of Binance's USDT-margined ether prints were exactly , and 41.6% of its COIN-margined ones. The flat ceiling in the figure is that default. Notice too that the stretches below it, where funding went negative and shorts paid longs, are long and shallow, and the excursions above are short and steep: through most of 2022 and 2023 the default was a ceiling the market kept falling away from, and in 2021 a floor it kept jumping off.

The third is that funding has edges at the other end as well. Venues cap and floor the rate, so in a violent move the print can stop measuring the premium at all and simply sit at the limit. At 16:00 UTC on 19 May 2021, the USDT-margined contract printed and the COIN-margined one , a suspiciously round number for a measurement and the kind a floor produces. The two contracts track the same asset on the same venue, are margined differently, and paid different rates at the same moment: their averages across the window are and a year. Which contract's funding you're reading is part of the reading.

The fourth is the clock. Binance and Bybit have long settled every eight hours, at 00:00, 08:00 and 16:00 UTC. Hyperliquid settles every hour, paying one eighth of an eight-hour rate each time, so its prints are an eighth of the size for the same pressure. And since 2025 even a single venue's clock moves: Binance switches a contract to hourly settlement once a print reaches its cap or floor, and moves it back out after a stretch of quiet. A print only means something next to its interval. To compare prints across venues or across time, put them all on a year:

with the print's interval in hours, read from the time since the previous print rather than assumed. The default every eight hours comes to a year, which is why "funding is about eleven percent" is the quiet state of a perpetual, with nothing bullish about it.

Aggregators complicate this further by averaging funding across venues, usually weighted by each venue's open interest. That blends contracts on different clocks, with different caps and different margin currencies, and, if a venue double-counted its open interest, gives that venue twice its weight. Wherever possible, read one contract's own prints.

#Basis and Premium

Part 5 defines the basis as the future's price minus spot, , and gives its annualised form. That definition hides an assumption that's easy to lose when the two prices come from two charts: and have to be read at the same moment. This is where my own chart comes in.

On 1 December 2023 I was asked whether the CME bitcoin premium held any edge, and I replied with an ether chart: CME's continuous ether future, ETH1!, on top, the ether index at the bottom, and between them a pane of the difference in their daily closes, which is what I meant by "premium delta" in the reply. Here it is reproduced as drawn, with the index swapped for Coinbase's daily close, since TradingView's index can't be exported.

It's a busy line. It swings between about and in a day, it spends barely more time above zero than below (54% of days), and its widest swings sit on the most violent days in the window. Futures on a coin with a cost of carry should mostly trade above spot, and by a margin that drifts rather than leaps, so something in it is wrong, and it took me until planning this post to see what.

The two closes aren't at the same time. CME's daily close for ETH1! is its settlement price, which CME sets from the trades in the last minute before 15:00 Chicago time, which is 20:00 UTC in summer and 21:00 in winter. The index's daily close is at 00:00 UTC. My pane was subtracting a price from mid-afternoon in Chicago from a price three or four hours later, and on any day ether moved in those hours, the move went straight into the "premium". Read spot at the moment CME settles instead, from hourly data, and the line changes character:

Its standard deviation falls from 41.5 to 13.7, it sits above zero on 67% of days, and the biggest spike of the original, on 12 May 2021, turns out to be . Almost all of what I was reading as premium was ether moving between two closes.

The reply itself drew a reading from that pane's volatility, and whether that reading survives the correction is a question for the Tool that uses it, where it can be run on the aligned series and judged. The lesson for reading data is the general one. Whenever you subtract one market's price from another's, check that the two prices are from the same moment. Daily closes on different venues, sessions and calendars almost never are. CME also closes for an hour every afternoon and for the weekend, while spot trades on, which is why its chart carries the gaps Daily vs Weekly Market Closes covers.

Then the unit. A gap of 100 dollars is a 2.5% basis when ether is at 4,000 and a 10% basis at 1,000, so a premium in dollars grows and shrinks with the price whatever the market thinks of carry. Divide by spot:

As a share of spot, the aligned premium shows something the dollar version buried. From September to the end of 2022 CME's ether future spent most days below spot, in backwardation: 81% of September's sessions and 94% of those from FTX to the year's end. The dollar version had it buried under the timing noise.

Two more traps live in a continuous series like ETH1!. It's stitched from one monthly contract to the next, so its premium jumps at each roll as the series switches from a contract about to expire, whose price has converged on spot, to one with a month to run. And because it hides which contract you're looking at, you can't annualise its basis with Part 5's formula without knowing the days to expiry. For a basis you mean to annualise, read a named contract.

#Aggressor Flow

Every trade has a buyer and a seller, but one of them arrived second. A trade happens when an order crosses the spread and meets one that was resting on the book: a buyer lifting an offer, or a seller hitting a bid. The one that crossed is the aggressor, or taker, and the resting one is the maker. That's the only honest sense in which a trade "was a buy" or "was a sell", and it's the directional measure the volume post promised was coming.

Binance publishes it in every candle: alongside the volume sits the taker-buy volume , so taker sells are , and the difference is

the taker delta in the venue's units and the taker-buy share.

The share is the one people quote, and it's the less useful of the two. On 19 May 2021, the day Binance's ether perpetual opened at 3,377, traded as low as 1,400.73 and closed at 2,443, its taker-buy share for the day was 0.483, and its spot market 0.485. A day of near-panic, and the number barely left a half. That's normal. Market makers, arbitrage desks and anyone splitting a large order cross the spread in both directions all day long, so the share sits near the middle almost regardless of what's happening. What carries the information is the delta, in units, and the timeframe it's read on.

In the 12:00 UTC hour on 19 May the perpetual's sellers crossed the spread for 83,345 ether more than its buyers, on a share of 0.467, and spot's for 54,758 more, on 0.437. The shares moved a few points; the deltas were the largest of the five days around them. Over the whole day the perpetual's delta was ether and spot's , so both venues were sold aggressively, the perpetual on more than twice spot's volume.

Comparing those two series, perpetual against spot, to see which one is leading, is Tool 02's job. Here it's enough to know what they count: aggression rather than participation, in the venue's own units, and only on the venue that reported it. The volume post's rules about units and venues carry straight over. A taker delta in ether from 2021 and one from 2023 aren't on the same scale of money, and Binance's flow says nothing about Coinbase's.

#Liquidations

Part 5 covers the liquidation engine: when a position's losses reach its maintenance margin, the exchange takes it over and closes it with a market order. The data is the record of those closes, and it has three problems: a sampled feed, labels that point different ways, and pictures of liquidations that haven't happened.

The feed first. Binance's public liquidation stream doesn't report every liquidation. Its documentation says that for each symbol only the latest liquidation order within 1000 ms is pushed as a snapshot, and if none happens in that second, nothing is. Coin Metrics dates the change to 27 April 2021; before it, Binance reported every one. In a quiet market that loses almost nothing. In a cascade, when hundreds of positions can go in the same second, it loses almost everything.

The first of the three dates on this post's charts, 19 May 2021, came three weeks after that change, so any liquidation total for it built from Binance's public feed is a sample, and the October 2025 crash that Part 5 mentions was argued over for exactly this reason. Totals compiled from public feeds are a floor.

Then the labels. A long position is closed by selling, so a long liquidation arrives on the tape as a sell order. Most venues report the side of that closing order, so "sell" there means "a long was liquidated"; Coin Metrics notes that Bybit is the exception in its coverage and reports the side of the position instead. Before adding liquidations across venues, check which way each one labels them, or you'll count longs as shorts.

And the pictures. A liquidation heatmap, the kind that shows bright bands of "liquidity" above and below price, isn't a record of anything. Exchanges don't publish their users' liquidation prices, so a heatmap is estimated from price, volume and assumed leverage levels, and CoinGlass's own explanation calls its scale a relative intensity and says actual liquidations may be lower than shown. It can be a useful picture of where leverage might sit. It's a model, and it should be read as one: Part 5's line that the chart shows liquidation levels is about the levels that exist, which no one outside the exchange can see.

#How This Series Reads Derivatives Data

The working rules, written plainly, for the Tools to link to.

1. Every long has a short. Open interest counts contracts, each with one long and one short. Never read it, or anything derived from it, as a count of one side (Open Interest).
2. Check what open interest counts. Contracts or sides of contracts (Bybit counted both sides until 11 June 2026), and on which venue (Open Interest).
3. Keep open interest in the unit that answers the question. Coins for positions opened or closed, dollars for money at risk: . Multiply contract counts by contract size before adding markets (The Unit Moves Under You).
4. Positions are data, reasons are inference. The CFTC's weekly report shows who holds each side of CME's futures, three days late. Why they hold it is a hypothesis (Who's on Each Side).
5. Long/short ratios count accounts. They're a headcount on one venue; the money behind every contract balances whatever the heads say (Long/Short Ratios).
6. Funding is a price with a blind band. A print of covers any premium from to ; a print at the cap or floor has stopped measuring (Funding).
7. Annualise funding at its own interval. , with read from the prints. About a year is the default state and carries no signal on its own. Read one contract's prints over an aggregate (Funding).
8. Subtract prices from the same moment. A basis or premium is with both read at once; for CME, at its 15:00 Chicago settlement (Basis and Premium).
9. Read basis as a share of spot. , and annualise only a named contract with known days to expiry; a continuous series jumps at its rolls (Basis and Premium).
10. Read aggression in units. The taker delta carries the information; the share sits near a half almost always. Daily bars average it away (Aggressor Flow).
11. Liquidation totals are a floor. Public feeds sample (Binance's pushes one order per symbol per second); check each venue's side convention before adding them (Liquidations).
12. Heatmaps are models. Estimated from price, volume and assumed leverage, and read as a model of where leverage might sit (Liquidations).

The shadow engine in Part 5 runs on contracts, and every one of them has two owners. Most of what goes wrong when people read its numbers comes from forgetting the second one, or from reading a price, a sample or a model as if it were a count. Hold on to the pair, check the clock and the unit, and the numbers start telling you what they were built to, which is usually less than people claim and occasionally a great deal more.