How Exchanges Build the Index Price: The Basket Behind Your Liquidation

Why the same position has two different liquidation prices

Open the same 10x long on BTC at two venues, with identical size, identical margin and identical maintenance requirements, and the two liquidation prices will not match. Not because either venue calculated the formula wrong, and not because one is quoting a stale number. They differ because the two engines are watching two different prices, each assembled from a different set of spot exchanges under a different set of rules.

This article is about how that number gets built. It assumes you already know that liquidation runs off the mark price rather than the last trade — crypto liquidations explained covers that, and perpetual futures explained covers the three-price framing — and asks the question those articles leave open: the index is "a volume-weighted average of major spot exchanges", but which exchanges, weighted how, and what happens when one of them stops behaving?

The answer is not a technicality. It is the difference between a liquidation price you can reason about and one you cannot.

Who is actually in the basket

Binance publishes its index constituents through a live endpoint, so the basket can be read directly rather than inferred. For BTCUSDT it contained eight venues:

Two things in that table are worth sitting with.

Binance's own spot market is the largest single input, at 43.5%. The venue that holds your position is also the venue supplying nearly half of the price that decides when to close it. That is not a scandal — Binance genuinely is the deepest BTC spot book, and excluding it would make the index less representative. Deribit does the same thing on purpose and says why: "If Deribit's own spot market is liquid enough, it is included in the index to be resilient against connectivity loss." A venue that cannot reach the outside world can still price its own book.

But it does mean the manipulation-resistance argument for index pricing is weaker than the phrase "basket of exchanges" suggests. An index diversifies away a wick on one small venue. It cannot diversify away a dislocation on the venue holding 43.5% of the weight — and October 2025, below, is what that looks like when it happens.

The pool Binance draws from is wider than any one basket, and the page is explicit that it is not a fixed list. Its methodology states that the index "includes prices from a broad range of exchanges, such as Binance, KuCoin, OKX, HitBTC, Gate.io, MEXC, Coinbase, Kraken, Bitfinex, Bybit, PancakeSwap (BNB Chain), Uniswap (Ethereum), Raydium (Solana) and Aster" — "such as", not "consisting of". None of Kraken, Bitfinex, HitBTC or any of the on-chain venues appeared in the live BTCUSDT basket above.

The page explains part of that gap directly: the on-chain constituents "will be available in contracts which are listed from 2025-02-10 onwards, subject to availability and price feed stability". BTCUSDT long predates that date, so it does not get them. And Binance "reserves the right to change the constituents of the Price Index from time to time without notice."

That is the practical lesson of this section. A venue's methodology page tells you the universe; only the per-symbol constituent list tells you what is actually pricing your position, and it can change without an announcement.

Other venues describe their baskets rather than enumerating them live. Bybit builds from "the top six (6) Spot trading pairs on the major Spot exchanges by trading volume". OKX requires "at least three" sources. Deribit says the larger the market the more constituents it includes, and that "the BTC/USD and ETH/USD indices typically have five or more constituents".

Three different theories of weighting

Once the constituents are chosen, they have to be combined, and the three published approaches optimize for different things.

Fixed preset weights (Binance, OKX). Weights are set by the venue and held until it changes them. Binance's BTCUSDT weights above are every one of them an exact multiple of 1/46 — 20/46 for Binance, 6/46 each for OKX and Coinbase, 3/46 each for the four mid-tier venues, and 2/46 for Gate.io. That is an allocation someone sat down and chose, not one the market produced. The advantage is that the weighting cannot be moved by a participant; the cost is that it goes stale as real liquidity shifts between venues.

Volume weighting, recomputed continuously (Bybit). Bybit weights each constituent by its 24-hour trading volume share, and states that "the index weights are updated hourly". This tracks liquidity as it actually moves. It also means the weighting is an output of market activity, which is a surface that can be pushed: a venue that manufactures volume earns index weight.

Risk-assigned weights (Deribit). Each constituent is "assigned a weight between 0 and 100%", set by a risk team against liquidity criteria that the documentation states are "assessed at the discretion of Deribit's risk management team". Deribit adds one structural note the others do not: "If Deribit's own spot market is liquid enough, it is included in the index to be resilient against connectivity loss."

None of these is obviously correct. They trade manipulation-resistance against responsiveness in different proportions, and a trader's only real move is to know which one is pricing their position.

What happens when a constituent breaks

This is where the designs genuinely diverge, and it is the part almost no article covers. Every venue has to answer the same question — one constituent is printing a price nothing else agrees with, now what? — and they answer it in two fundamentally different ways.

Camp one: cap the outlier toward the median

Binance and Deribit both keep the deviating constituent in the basket but refuse to let it move the index past a bound.

Binance's rule: "If the latest price of a specific exchange deviates by more than 3% from the median price of all sources, the value will be immediately capped at either 1.03 times or 0.97 times the median price, depending on whether the deviation is above or below the median."

Deribit applies the same idea with a band six times tighter, describing its method as a capped median. Its documentation works a full example: given sources at 100, 101 and 120 with weights of 25%, 50% and 25%, the initial median is 101, and each price is clamped to within 0.5% of it. The outlier at 120 contributes 101.505, not 120. The final index is 101 — the wildly dislocated third source moved it by nothing at all.

Run the same three prices without the cap and the weighted average is 105.50. With it, 101.00 — exactly the median. A source printing a 19% premium was worth 4.5 points of index before the rule and zero after it. That is the clearest one-line statement of what an index rule is for.

Camp two: throw the outlier out

Bybit takes the opposite route. Beyond the threshold the component is removed rather than clamped: "If the Spot price of any component trading platform diverges by more than 5% from the median of all Spot price sources, the system will temporarily exclude the respective component from index price calculation", with its weight "gradually reduced using a smoothing algorithm and redistributed among the remaining non-excluded components" until its price comes back within range.

The headline 5% is not the number that matters for most readers. Bybit adjusts it per asset, and for the two assets most people trade it is five times tighter: "BTC & ETH is 1%, Gold and Silver pairs such as XAUUSDT and XAGUSDT is 3%".

So the practical spread across four venues, for BTC, runs from Deribit's 0.5% band to Bybit's 1% exclusion trigger to Binance's 3% cap — with OKX declining to publish a number at all, stating only that a source deviating "beyond a defined threshhold from the median price of the rest of the price sources" will be "adjusted toward the median accordindly".

The same dislocation, three different indices

The cleanest way to see what those rules cost is to hold everything else constant and change only the rule. Take five constituents at 20% each, four of them agreeing and one dislocated 7.5% below the rest:

Work it through. The four healthy venues print 62,700, 62,710, 62,720 and 62,730; the fifth prints 58,000. The median of all five is 62,710, and every rule measures from there.

Deribit's 0.5% band clamps the outlier to 62,396.45, so it still contributes a fifth of the weight — just at the band edge. Binance's 3% band is six times wider, so the outlier contributes 60,828.70 and drags the average further down. Bybit's rule sends the outlier out of the basket entirely and splits its weight across the four survivors, so the index becomes the clean average of the venues that agreed.

The result is the opposite of what the thresholds suggest. The widest tolerance produces the most dislocated index, because a loose cap admits more of the outlier; the tightest produces one closer to the healthy consensus; and dropping produces the index furthest from the outlier of all. A trader reading "3%" as more permissive and therefore safer has it backwards — permissive here means the bad price gets more of a vote.

Putting a position underneath it

The spread across the three indices is 377.26, or 0.60% of price. Now give that spread something to act on.

Take a 20x long opened at $65,445, on a venue charging a 0.5% maintenance margin rate. Its liquidation price is entry × (1 − 1/20 + 0.005) = $62,500. The market has since fallen 4.2%, so the position is alive but close to the line — which is the ordinary condition of a leveraged book during the kind of stress that dislocates a constituent in the first place.

Then one constituent prints 58,000:

Same position, same size, same leverage, same maintenance rate, same five prices, same second. On two of the three rules the trader still holds the position. On the third the risk engine has already closed it and taken the margin. Nothing separates those outcomes except which tolerance the venue wrote into its index methodology — and the one that closed the position is the one with the widest published band.

When the feed simply stops

A constituent that goes quiet is a different failure from one that dislocates, and each venue sets its own patience:

• Binance — "If Binance is unable to access data from an exchange or the exchange has not updated its trading data within the last five minutes, the weight of that exchange will be set to zero."
• Bybit — "if no Spot trading pair has been traded on the exchange for more than 15 minutes, the trading pair will be excluded from both the index price calculation and the median calculation pool."
• Deribit — "If a constituent does not update at least every 3 minutes, that constituent is removed from the index."
• OKX — sources "that underwent system maintenance or didn't update their latest price during a specified time period won't be taken into calculation", without publishing the period.

Bybit adds a subtler mechanism for a book that is technically alive but too thin to trust. Rather than using the last trade, it derives a price from resting liquidity: Ob Price = (AskPrice1 × BidVolume1 + BidPrice1 × AskVolume1) ÷ (BidVolume1 + AskVolume1). As Bybit puts it, "when one side of the orderbook becomes significantly thinner, the weighted price tilts toward it, making it easier for the price to move in that direction."

When the whole basket fails at once

The tail cases are where the designs reveal what they actually value.

Bybit degrades, then goes self-referential. If every constituent deviates beyond the threshold, "the weights of the excluded exchanges will be smoothly redistributed to the remaining included exchanges until only one exchange is left (components that deviate from the median earlier will be excluded first)". And if no reasonable spot price can be obtained anywhere — including from Bybit itself — "the index price will be calculated from the last traded price of the Perpetual Contract", smoothed as Index at Tn = α × Target Price at Tn + (1−α) × Index at Tn−1 with α defaulting to 0.1818. At that point the index references the contract it is supposed to be pricing.

Deribit halts instead. "Deribit requires that each index used for settlement has at least one active constituent. If an index has no constituents, then all related derivatives halt until the underlying issue is resolved." It layers circuit breakers on top: "If an index moves more than 10% in between index ticks (1 second), the related derivatives are halted." And when only two sources survive, "the maximum allowed spread between them is 2.5%".

OKX degrades quietly. With two valid sources "they will be weighted equally"; with one, "it will be taken as the index price" — a single venue's last trade becomes the number your position is marked against, with no halt and no disclosure threshold.

That is the choice in one line: keep trading on a degraded price, or stop trading. Bybit and OKX keep going; Deribit stops.

From index to mark, in one pass

The mark price is what your position is actually valued at, and every venue derives it from the index — this article assumes the why, which crypto liquidations explained and funding rates explained already cover. What is worth putting side by side is that the derivations are not the same shape.

Binance's is the only median-of-three: Mark Price = Median (Price 1, Price 2, Contract Price), where Price 1 = Price Index * (1 + Last Funding Rate * (Time Until Next Funding / Funding Period)) and Price 2 = Price Index + Moving Average (30 seconds basis). Taking a median of three candidates means a single misbehaving input — including the contract's own price — is outvoted rather than averaged in.

OKX states the simpler form directly: "Mark price = spot index price + basis moving average", and confirms the consequence plainly — "Forced liquidations take place based on the mark price instead of the last traded price."

Deribit smooths the premium rather than the price, applying an EMA to "Futures price − Index price" and adding it back, so that "the mark price remains anchored to the index", then applies "final dynamic bandwidth limits" around the index. Its documentation attaches a warning most venues do not: "the mark price should not be relied upon for making trading decisions… sudden price movements often result in the mark price falling outside the bid-ask spread."

October 2025: when the broken constituent was the venue itself

The clearest evidence that a basket is a live risk parameter is what happens after one fails.

Between 2025-10-10 21:36 and 22:16 UTC, three assets — USDE, BNSOL and WBETH — depegged on Binance's own spot market. Binance compensated affected users by reference to "the difference between market price at 2025-10-11 00:00 (UTC) and their respective liquidation price".

The structural fix arrived three days later, and it is the part worth reading closely. Binance Margin changed what those indices were made of. BNSOL's price index moved from Binance: BNSOLUSDT (30%) + Binance_cross: BNSOLSOL*SOLUSDT (70%) to Binance_cross: BNSOLSOL*SOLUSDT (100%), with the conversion ratio based on the official SOL redemption rate. WBETH moved from Binance: WBETHUSDT (20%) + Binance_cross: WBETHETH*ETHUSDT (80%) to Binance_cross: WBETHETH*ETHUSDT (100%). The stated reason: "to minimise the risk of depegging".

Read what that change actually did. The old baskets included Binance's own thin quote market for the wrapped asset, at 20–30% weight. When that one book dislocated, it dragged the index — and the index was pricing collateral. The fix removed the direct quote entirely and replaced it with the wrapped asset's ratio against its underlying, anchored to a redemption rate. The dislocated constituent was not an outside exchange. It was the venue's own book, inside its own index, pricing its own collateral.

Bybit's documentation shows a different answer to the same class of problem, arrived at independently: for BBSOL, CMETH, METH and USDE it now incorporates "the redemption price as part of the index price reference" and enforces a floor — Minimum Index Price = Redemption Price × 0.95 for the first three, and × 0.9 for USDE.

What to check before you size a position

Conclusion

The index price is not an observation of what BTC is worth. It is a construction: a venue picks a set of exchanges, assigns them weights on one of three incompatible theories, and writes rules for what to do when one of them dislocates, goes quiet, or when all of them fail together. Every one of those choices is a parameter, and no two venues have chosen the same values.

The differences are not marginal. The deviation threshold for BTC spans an order of magnitude across four venues, from Deribit's 0.5% band to Binance's 3% cap — and as the worked example shows, the wider band is the one that lets more of a dislocation through, which is the reverse of how a threshold intuitively reads. Two venues cap outliers and one drops them, pushing the index in opposite directions under identical stress. Patience with a dead feed ranges from three minutes to fifteen. And when the whole basket fails, one venue halts trading while another starts pricing the contract against itself.

That is the answer to the question this started with, and the worked example is the shape of it: one 20x long liquidating at $62,500, one dislocated constituent, three published rules, and a position that survives on two of them and is closed on the third. The gap is a property of the venue's risk design, not noise.

It is also, in fairness, rare. On an ordinary day these four indices agree to within a few basis points and none of the rules above does any visible work. They separate when a constituent dislocates — and constituents dislocate during precisely the conditions that have already pushed leveraged positions to the edge of their maintenance margin. That correlation is the whole risk: the rules are invisible until the moment they are decisive.

The practical discipline is narrow. Read the constituent list for the specific symbol you are trading, not the venue's general methodology page. Know whether your venue caps or drops, and at what threshold. Check how much of the index is supplied by the venue holding your position. And for anything wrapped, staked or synthetic, find out whether the index trusts a quote market or a redemption ratio — because October 2025 showed what the first one costs, and both major venues that got burned arrived at the same fix.

Frequently asked questions

Sources and further reading

• Binance — What Are Mark Price and Price Index in USDⓈ-Margined Futures?
• Binance — Query Index Price Constituents (API)
• Binance — Binance Margin Will Update the Price Index Components of BNSOL and WBETH
• Binance — Resolution of USDE, BNSOL, and WBETH Price Depeg
• Bybit — Index Price Calculation
• OKX — Spot index prices
• OKX — What are the index price and mark price of margined contracts?
• Deribit — Index Prices
• Deribit — Mark Prices

Related CoinBeaver articles: crypto liquidations explained for the liquidation formula this assumes, perpetual futures explained for the three-price framing, forced liquidation venues explained for the on-chain oracle analogue, funding rates explained for the mark-index premium, and how crypto leverage works for the position mechanics.