Reading public derivatives positions

Why some positions are public

On a centralized venue, a position is a row in a private database. The exchange can see it, as can regulators and courts with authority, but outsiders cannot. On a venue whose accounts live in chain state, positions are readable by anyone because every validator or node needs that state in order to compute margin. Publicity is a by-product of design, and rarely a feature chosen for analysts.

Access can be limited by region. For example, the dYdX indexer and Binance API returned regional blocks to some requests made while this lesson was prepared.

Traders know their positions are visible, and this creates a selection effect. Someone who wants privacy, whether to protect a strategy or simply by preference, can use a venue without public records, or split activity across addresses. So a public record is not a random sample of crypto derivatives positions. Chitra’s review notes that Binance and Bybit together processed nearly $58.5 trillion in perpetual notional in 2024, on private books. Any claim about what “whales” in general are doing, drawn only from a public venue, generalizes from an unrepresentative part of the market.

Anatomy of a position record

The Hyperliquid clearinghouseState response is a useful teaching case, since it is free, well documented and detailed. Other venues have similar fields under other names.

Field names from Hyperliquid’s API documentation, read November 2026. Comments are this lesson’s.

On funding’s sign, the sample rows support a reading in which a positive cumFunding means the position paid. A short with funding at a positive rate showed a negative sinceOpen, and a long showed a positive one. That is an inference from a few rows, not a stated definition. Always confirm the convention on a case you can verify before building on it. userFills and userFunding give fuller history, including closed positions, which a current-state snapshot cannot show at all. Lesson 6’s warnings about cost basis and survivorship apply here without change.

Addresses are not entities

Lesson 7 covered attribution. Derivatives add a sharper version of the problem, because a position’s meaning depends on what offsets it. The opening JELLY case involved offsetting accounts. HLP shows the same structure in the open. The venue publishes a parent vault with child addresses, so it can be studied without any attribution guesswork.

A related trap involves multiple accounts and subaccounts. Some venues support several subaccounts per wallet, as dYdX does with its subaccount number, and a trader can use separate wallets with no on-chain link, or link them only through an exchange withdrawal. Standard clustering heuristics, such as a shared funding source, can suggest a link. They can also create false ones, for example where many users withdraw from one exchange’s hot wallet. Timing correlations between accounts are weak evidence on their own. So state whether an entity claim comes from a public official statement, an established public attribution, or your own heuristic. Treat the last of these as a hypothesis, and do not attach a real person’s name to it.

Hedges you cannot see

A public record shows one venue’s leg of what may be a multi-leg position. A large public short may be any of the following.

Readings are not exhaustive and are not mutually exclusive.

He et al. model a market in which end users demand perpetuals for leveraged exposure while arbitrageurs take the other side and hedge in spot, so some shorts on any book may be hedges by design. Their evidence concerns price deviations and capital constraints, however, and not a measured split of open positions. No public dataset read for this lesson gives that split. The honest default is to describe a public position as what it is, a position on one venue, rather than as a view.

Modeled heatmaps and real liquidation prices

Two different objects get called liquidation data, and they often get confused.

A simple construction shows what a heatmap assumes and why it can drift from the truth. Take an estimate of position opening at price P₀ and assign it leverage buckets. Each bucket produces a level:

Models therefore fail in predictable ways. OI does not reveal which side opened. Leverage is assumed, and on cross-margin accounts it differs from a label. Positions change: partial closes, margin added, profit accumulating. Models usually cover only some venues. A zone that looks large may also never be tested, and a heatmap by itself does not show that price is attracted to it. That last claim, widely repeated, needs a test with fixed rules, a benchmark and base rates, by the methods of Lesson 9. Evidence here is thin. Results also depend on the model and the provider, and a provider’s own heatmap may offer different versions.

Observed prices have limits too. A cross-margin level assumes all other positions stay constant, as the $1.04 billion BTC figure shows. On a venue with partial liquidation, a large position may not close all at once. For example, Hyperliquid sends 20% of liquidatable positions above 100,000 USDC to the book. Actual closure depends on the mark price, not the last trade. And margin can be added at any moment before the level is reached, so a level that looks near may move away by the time the price arrives, without any trade. Reported liquidation streams, on the other hand, measure what did happen, but venue conventions differ. Coin Metrics notes that OKEx reported the bankruptcy price rather than the fill price, that BitMEX reported the size still open rather than the amount filled, and that Binance’s stream shows at most one update per market per second.

Market makers in position data

Large public books often belong to market makers, whose positions describe a business, not a forecast. A market maker’s inventory comes from filling other traders’ orders. It tends to be hedged elsewhere, and it may reverse within minutes.

Some features suggest this activity, though none proves it: many positions across many assets, both long and short; gross notional much larger than net; low effective leverage; very many fills on both sides with small average size; and a repeated tendency to revert toward flat. The HLP example above has every one of the position features, and the venue describes it as a market-making vault. For an unlabeled address, those features justify only a hypothesis, since a systematic hedge fund or an arbitrage bot might look similar.

Exploratory measures with no accepted thresholds. Applied to one address, each inherits the entity problem above.

Drawing conclusions honestly

A public position supports only a narrow class of claims. The discipline is to write the claim in a form the data can actually support, and to state the limits in the same sentence.

Addresses and figures in this table are examples.

There is also a standing ethical point. A public address is not a public person. Publishing a claim that links a position to a named individual can cause real harm if it is wrong, and it invites targeting even if it is right. Use only attributions that come from public official statements or established public sources, name the source, and say how strong the link is. Lesson 7 gives the fuller argument.

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.