Source: https://tesseralytics.dev/blog/oi-composition

[← Blog](https://tesseralytics.dev/blog)

Research 13 November 2026

# Open interest has no direction. Its composition does.

A perpetual’s open interest rises by the same amount whether new longs arrived or shorts capitulated, so direction has nowhere to live in the level. Measured over 13 months and 127 Hyperliquid markets: who arrives carries a sign in both crypto and HIP-3, and pooling the two manufactures a result.

Open interest is the most quoted number in derivatives and one of the least informative. When a market's OI jumps 5% in an hour, two very different things are wearing the same number. Someone new arrived and took a position. Or a crowd of shorts gave up and closed, and longs replaced them. The first is fresh conviction. The second is the tail end of a squeeze. Both leave open interest higher by exactly the same amount.

That is not a gap in the data. It is arithmetic. A perpetual's ledger is two-sided and complete: every contract has one long and one short, so a one-sided open interest figure counts each contract exactly once, and it moves by the same amount whether the position that opened was long or short. **Direction has nowhere to live in the level.**

It does live somewhere: in the *composition* of the accounts behind the change. This post measures that composition over 13 months of Hyperliquid data (1,111,069 coin-hours across 127 markets), and then does the thing that turned out to matter most: it measures two different kinds of market separately, because pooling them invented a result that was not there.

## What a perpetual's ledger contains

The two-sided ledger leaves a fingerprint. Every hour, each account is either long, short, or flat, and the transitions between those states are countable. The `gold_positioning_1h` tile publishes them as six counts per coin per hour:

-   **Arrivals**: `n_new_longs`, `n_new_shorts`, the accounts that went from flat to a position.
-   **Exits**: `n_closed_longs`, `n_closed_shorts`, the accounts that went from a position to flat.
-   **Flips**: `n_flipped_l2s`, `n_flipped_s2l`, the accounts that crossed zero, long to short or short to long, inside the hour.

Two warnings, both of which cost people real money when they skip them. A flip is counted twice: once as a close and once as a flip, so the exit side is a superset of the flips, not a third disjoint bucket. All six count accounts, not size. A wallet adding a hundred times its position is invisible in every count. Size lives in the taker columns (`oi_opened`, `oi_closed`), which cover the taker side only. The maker side is the residual, and it is not observable from fills.

The identities that make the decomposition trustworthy are testable, so they were tested rather than assumed. On all 855,347 consecutive-hour rows of the crypto panel and 253,618 of the HIP-3 one, `Δn_long ≡ n_new_longs − n_closed_longs` exactly, same for shorts, with zero residual on every row.

And the gross is much bigger than the net. In the median crypto market, accounts make **3.0 transitions** for every one position of net change in the position count, and taker volume opened plus closed runs at **6.2×** the absolute change in open interest (the HIP-3 panel: 3.0× and 6.0×). An OI move of two tenths of a percent is not a quiet hour. It is the residue of a lot of people opening and closing.

## How the composition gets measured

The test is the same one used for any factor: rank every market by the field at some hour, rank them again by the return over the following hours, and correlate the two rankings. That number, the **information coefficient** or IC, says whether the field told you anything about which markets would outperform. It is a measurement, not a promise, and a field with a small IC is information too: it says there is nothing there.

Entries are at `H+2h`, not at the close of the hour being measured. That delay is not a modelling choice: a bar's close does not exist until the bar is over, the venue's open-interest snapshot lands after the boundary, and the row takes a pipeline cycle to become readable. Two hours is the first entry anyone can actually take; the [previous post](https://tesseralytics.dev/blog/the-two-hour-lag-discount) measures what that delay costs. Every number below is at that entry, so none of them is a backtest of a price you could not have got.

The universe is liquidity-floored (median hourly traded notional of at least `$10,000` and at least 50% of the best-covered coin's hours), because on a thin coin the counts are single digits and a flip *share* is a ratio of noise. What that leaves, in each population:

| panel | markets | coin-hours | median notional / h | median accounts |
| --- | --- | --- | --- | --- |
| Crypto perpetuals | 92 | 855,496 | $43,036 | 1,428 |
| HIP-3 (xyz:\*) | 35 | 255,573 | $178,947 | 893 |

The HIP-3 panel is wealthier per market and thinner per account: four times the notional, three-fifths the holders. Both panels clear the floor comfortably, so liquidity cannot separate the two populations, which is why the split has to be explicit.

Then the fields, and the sign convention without which the tables are unreadable:

| field | what it is | a positive value means |
| --- | --- | --- |
| doi\_rel | the level: ΔOI as a share of the market | open interest rose |
| n\_new\_tilt | (new longs − new shorts) / arrivals | the marginal entrant is long |
| n\_close\_tilt | (closed shorts − closed longs) / exits | shorts are the ones leaving |
| n\_flip\_tilt | (s→l flips − l→s flips) / flips | shorts are being squeezed out |
| flip\_share | flips / transitions; direction-free | more of the activity was a reversal |
| churn | arrivals + exits / unique accounts | the holder base turned over |

## The level ranks nothing in either market

Start with the number everyone quotes. Rank markets by how much their open interest changed this hour, and ask which of them outperform over the next one, six and twenty-four hours.

| panel | h=1 | h=6 | h=24 | t @ 6h |
| --- | --- | --- | --- | --- |
| Crypto | −0.0006 | −0.0005 | 0.0008 | −0.35 |
| HIP-3 | 0.0021 | 0.0006 | 0.0045 | 0.22 |

Rank IC at a 2-hour entry. Crypto: 92 coins, 855,496 coin-hours. HIP-3: 35 markets, 255,573 coin-hours.

Neither panel shows anything at any horizon: the crypto |IC| never exceeds 0.001, and the HIP-3 level tops out at 0.0045 at twenty-four hours, on a six-hour t of 0.22, which is what a cross-section of 35 markets with no effect looks like. For scale, a field has to clear roughly 0.005 here to be worth a second look, and the strongest factors in the previous post peak near 0.05.

This is the first durable result of the study, and it is a negative one: the size of an open-interest change is direction-blind information, exactly as the ledger arithmetic said it would be. If you have ever ranked markets by OI growth, you were ranking noise.

## Everything measured, per panel

The whole family, both populations, at the entry a customer can take. Nothing here is pooled, and the two tables are deliberately not averaged into one. The level row is in the table above, so both families start at the arrival tilt.

Crypto perpetuals: 92 markets, 855,496 coin-hours

| field | kind | h=1 | h=6 | h=24 | t @ 6h | BH |
| --- | --- | --- | --- | --- | --- | --- |
| n\_new\_tilt arrival tilt | arrival tilt | 0.0028 | 0.0035 | 0.0046 | 2.59 | yes |
| n\_close\_tilt exit tilt | exit tilt | −0.0009 | −0.0014 | −0.0021 | −1.02 |  |
| n\_flip\_tilt flip tilt | flip tilt | −0.0004 | −0.0002 | 0.0000 | −0.22 |  |
| flip\_share flip share | shape | −0.0017 | −0.0036 | −0.0014 | −1.90 |  |
| churn | shape | −0.0138 | −0.0235 | −0.0286 | −8.87 | yes |
| oi\_flow\_rel taker flow | taker size | 0.0000 | 0.0006 | 0.0015 | 0.44 |  |
| taker\_flip\_oi\_share taker flip share | taker size | −0.0069 | −0.0124 | −0.0200 | −5.99 | yes |

HIP-3 (xyz:\*): 35 markets, 255,573 coin-hours

| field | kind | h=1 | h=6 | h=24 | t @ 6h | BH |
| --- | --- | --- | --- | --- | --- | --- |
| n\_new\_tilt arrival tilt | arrival tilt | 0.0027 | 0.0088 | −0.0011 | 3.04 | yes |
| n\_close\_tilt exit tilt | exit tilt | −0.0040 | 0.0003 | 0.0019 | 0.12 |  |
| n\_flip\_tilt flip tilt | flip tilt | −0.0013 | 0.0023 | −0.0004 | 1.09 |  |
| flip\_share flip share | shape | 0.0012 | −0.0018 | −0.0035 | −0.49 |  |
| churn | shape | 0.0074 | 0.0191 | 0.0151 | 3.41 | yes |
| oi\_flow\_rel taker flow | taker size | 0.0040 | 0.0020 | 0.0046 | 0.75 |  |
| taker\_flip\_oi\_share taker flip share | taker size | 0.0006 | 0.0067 | 0.0068 | 1.40 |  |

Gross rank IC, 2-hour entry. BH is the family-wide Benjamini–Hochberg test at 5% over all 48 (field × horizon × lag) tests; 18 clear it in crypto and 5 in HIP-3. A smaller panel has less power, so fewer of its fields clear any bar; that is a fact about the panel, not the effect.

## Who arrives carries the sign

The composition does better. The field that survives in both panels is the **arrival tilt**: the share of the hour's arrivals that were longs, on a scale where zero means they split evenly.

In crypto it reads +0.0035 at six hours (t 2.59) and holds its sign at one and twenty-four. In HIP-3 it reads +0.0088 (t 3.04), which is *stronger* than the crypto reading on a panel with less than a third of the data. And it is not a repackaged level or a repackaged momentum: its cross-sectional rank correlation is −0.06 with the OI change in crypto and 0.00 in HIP-3, and 0.01 with last hour's return in both. It is also fast: an autocorrelation of 0.07 after one hour in crypto and 0.03 in HIP-3, so it is measuring this hour's arrivals, not a slow property of the market.

*When the money arriving in a market is one-sidedly long, that market tends to outperform other markets over the next day.* Not much, and not reliably enough to trade (more on the cost of that in a moment), but consistently, in two markets that share a ledger schema and little else. Two related fields do not survive: the exit tilt and the flip tilt are indistinguishable from zero in both panels (−0.0002, t −0.22 in crypto; +0.0023, t 1.09 in HIP-3).

There is one honest caveat, and it belongs in the headline rather than a footnote. Within the crypto panel the arrival tilt is **not** robust: it clears significance in the second half of the sample and in the liquid third of the universe, and vanishes (t 0.10) in the illiquid tail. The field that *is* robust inside crypto is `churn`, which holds the same sign with |t| above 3 in all four time-and-liquidity sub-panels. And `churn` is about to become the problem.

## Then the two markets disagreed about everything else

The same ledger, the same fields, the same measurement, run on two populations: crypto perpetuals, and HIP-3 markets. The `xyz:` tokens track equities and commodities, whose underlying cash market keeps office hours and closes at the weekend.

`churn` is the most robust field in crypto: busy hours underperform, at −0.0235 (t −8.87) at six hours. In HIP-3 it reads +0.0191 (t 3.41): **the opposite sign.** The taker flip share does the same thing, −0.0124 (t −5.99) against +0.0067 (t 1.40).

It gets sharper when you stop looking at the cross-section and start looking at single moves. Take every hour where a market's OI moved in its own top 30% for that market (a "real move"), and split the hours by the arrival tilt *that same hour*. Now you are comparing identical-size OI moves in the same direction, differing only in who did it. At 24 hours, the high-tilt-minus-low-tilt spread is:

Same-direction, same-size OI moves, split by the arrival tilt within the hour. Bars are the high-tilt-minus-low-tilt spread in forward return at 24 h, in basis points, within-hour demeaned, at a 2-hour entry. A bar's sign is the direction the tilt pays in.

Crypto perpetuals

build

+4.6 · t +0.64

unwind

+17.7 · t +3.01

HIP-3 (xyz:\*)

build

+35.9 · t +2.49

unwind

−12.9 · t −1.14

OI building OI unwinding not significant (|t| < 2)

Crypto pays when OI is *falling* and the accounts still arriving are longs (+17.7 bps, t 3.01). HIP-3 pays when OI is *rising* and the arrivals are longs (+35.9 bps, t 2.49). Each panel's live cell is the other's dead one: crypto's build cell is +4.6 bps (t 0.64), HIP-3's unwind cell −12.9 (t -1.14).

The event study says the same thing in a blunter form. A short squeeze (a spike in accounts flipping short-to-long, measured against matched non-spike hours in the same OI-move neighbourhood) is worth +15.8 bps at 24 hours in crypto (t 2.24) and −15.3 bps in HIP-3 (t -1.99). An event that pays one way in one market and the opposite way in another, at the same threshold, on the same measurement, describes the market structure rather than the event.

## Why the split is the method

The first version of this study ran one cross-section of 127 coins and reported the market split as an appendix. That was wrong in a way that produced findings rather than just errors.

An IC is a correlation of *ranks across a cross-section*. Put two populations with different levels of a field and different average returns into one ranking, and the gap between the populations becomes part of the correlation, which is what the statistic measures. Pooled over both panels, the arrival tilt's IC came out *higher* than in either panel separately, and the excess was a between-class component of +0.0021 with a t of 2.34, a "finding" made entirely of the fact that HIP-3 markets are richer per market, thinner per account, and listed later than crypto.

The same pooling also did quieter damage: it made `churn` look like it flipped sign between liquidity tiers (43% of the top liquidity tercile in the pooled run was HIP-3), turned a two-sided asymmetric split table into an apparent symmetric one, and averaged a +15.8 bps squeeze against a −15.3 one.

**A pooled rank IC over a heterogeneous universe manufactures signal.** It generalises far beyond this dataset: before you pool, ask whether the groups differ in level as well as in behaviour. Here they did, and the price was a result that did not exist.

What survives, measured separately and never averaged:

| claim | crypto | HIP-3 | verdict |
| --- | --- | --- | --- |
| ΔOI/OI ranks nothing | −0.0005 | 0.0006 | stable |
| flip tilt ranks nothing | −0.0002 | 0.0023 | stable |
| arrival tilt positive at 1–6 h | 0.0035 | 0.0088 | sign-stable, strength differs |
| churn is bad | −0.0235 | 0.0191 | opposite signs |
| taker flip share is bad | −0.0124 | 0.0067 | opposite signs |
| split table's live cell | +17.7 | +35.9 | opposite directions |
| short-squeeze spike | +15.8 | −15.3 | opposite signs |

IC at a six-hour horizon and 2-hour entry; splits and events at 24 hours, in basis points. Three claims survive, four do not.

## Why HIP-3 behaves differently: the underlying closes

A crypto perpetual's underlying *is* the same 24/7 book. An `xyz:` token tracks an equity or a commodity whose cash market is open 32.5 hours a week, so its holders carry an overnight and weekend gap with no crypto equivalent. That difference is visible in the data as a clock.

| session (ET) | hours | median |ΔOI/OI| | account transitions | holders | tilt IC @ 6h |
| --- | --- | --- | --- | --- | --- |
| cash session | 1,650 | 0.93% | 40 | 786 | +0.0158 · 2.35 |
| overnight | 3,960 | 0.36% | 22 | 791 | +0.0057 · 1.37 |
| weekend | 3,593 | 0.13% | 14 | 804 | +0.0092 · 1.97 |

HIP-3 panel only. US market holidays are not modelled, which dilutes the cash-session bucket slightly and cannot manufacture the difference between buckets.

Positioning in these markets moves **7× more per hour** when the underlying is trading than over the weekend: 0.93% against 0.13% of open interest, on roughly 3× the account transitions, while the number of holders barely moves (786, 791 and 804). That is what a market with a real underlying and a closed cash market looks like: same crowd, much more action when the thing it tracks is open.

And the composition's information follows the clock. The arrival tilt is +0.0158 (t 2.35) in the cash session against +0.0057 (t 1.37) overnight, **2.8× more informative**, and the cash session is the only bucket where it is significant at every horizon. The composition says most when the underlying actually trades.

## The dramatic event is the empty one

Capitulation is the story people tell about open interest: longs give up, flip short, OI falls, and the bottom is in. It is measurable: a trailing robust z-score of the flip count per market, against matched control hours on the same market at the same size of OI move.

It is a null. In crypto, long-capitulation spikes deliver −1.6, +1.8 and −3.5 bps at one, six and twenty-four hours; in HIP-3, −0.2, −2.7 and −6.9. Every one of the six cells sits inside |t| 1.5. The **control** event (a spike in plain *new longs*, with no flip involved) is the strongest row in the crypto panel: +11.6 bps at six hours, t 3.50. Arrivals again, in event form.

The label is earned. Flip counts do track aggressive flip volume more tightly than plain closes (a median rank correlation of +0.47 against +0.30, positive on every market in both panels), so a flip spike really is capitulation, not just an exit. It still buys nothing, because a flip *count* is 0.86 rank-correlated with total account churn: a capitulation spike is largely an activity spike, and activity carries no forward information here.

## Is any of this tradeable? No.

A rank IC is not a strategy. To become one the ranking has to survive its own turnover, and the arrival tilt is a fast field: 61% of the crypto cross-section changes tercile every hour. Charging a flat 10 bps per unit of turnover, the crypto tilt's top-minus-bottom spread is worth +2.7 bps gross at six hours against +10.0 bps of cost, for **a net of −7.3 bps**. Across both panels and all three horizons the arrival tilt's net lands between −13.1 to −3.4 bps: every one of them is negative.

The composition is a real measurement of who holds a market. It is not a tradable edge at this venue's costs. What it is good for is description: knowing whether an OI move was arrival or unwind is information about market structure, even when it is not information about the next day's price.

## What this does not say

-   No cost model beyond that gate. Returns are gross throughout; the cost column charges 10 bps per unit of turnover and nothing else. No spread, no impact, no borrow, no funding on the short leg.
-   One regime. 13 consecutive months, every statement cross-sectional and within-sample. Nothing here is a time-series claim.
-   Currently-listed markets only. Delisted coins are absent, so survivorship is not controlled for.
-   Counts are accounts, not size. A wallet adding 100× size is invisible in every count here; size lives in the taker columns only.
-   A snapshot grid is not a tape. A position opened and closed inside one hour does not exist in this data.
-   The tilt's sign is measured, not explained. "The marginal entrant being long predicts higher relative returns" is a fact about these venues and this sample. It is not evidence of informed flow: an uninformed crowd of small longs would produce the same IC, and nothing here distinguishes the two.
-   The HIP-3 panel is young. 5,000–9,200 hours per market against a full 9,503; the cohort is a recent listing wave, not 13 balanced months.

## Run it yourself, and what one month can show

Everything above is one script, and it is public: [github.com/tesseralytics/examples](https://github.com/tesseralytics/examples). It reads the positioning tile and one-minute closes, builds the eight fields, applies the floor once across the labelled universe, splits it, and prints the family and the cost gate; the session and event tables are not in it.

A free key covers the trailing 30 days, which is one closed month. That is enough to reproduce some of this and not the rest, and the honest split matters more than a chart a free reader cannot verify.

| what a one-month, free-key run can establish | reproducible? |
| --- | --- |
| the identities, the gross/net ratios, the flip-versus-taker correlations | exactly: properties of the ledger, not estimates |
| crypto `churn` and `taker_flip_oi_share` | yes, same sign, same order of magnitude |
| the arrival tilt in crypto | no: the effect is small and 720 hours cannot resolve it |
| the HIP-3 off-hours effect | no: 720 hours split three ways is not a session study |

```
export TESSERA_API_KEY="sk_..."   # free key, no card: https://tesseralytics.dev
git clone https://github.com/tesseralytics/examples
cd examples
```

Every field in the tables above is one line on the published columns:

```
import tessera

client = tessera.TesseraClient()          # reads $TESSERA_API_KEY
pos = client.read(
    "gold_positioning_1h",
    coin=["BTC"],
    month=tessera.MonthSpan("2026-10", "2026-10"),
    columns=[
        "time", "open_interest", "d_open_interest", "n_unique_users",
        "n_new_longs", "n_new_shorts", "n_closed_longs", "n_closed_shorts",
        "n_flipped_l2s", "n_flipped_s2l",
        "oi_opened", "oi_closed", "taker_flip_l2s_1h", "taker_flip_s2l_1h",
    ],
)
```

```
# From the raw columns, the decomposition the post measures.
prev = pos["open_interest"] - pos["d_open_interest"]
arrivals = pos["n_new_longs"] + pos["n_new_shorts"]
exits = pos["n_closed_longs"] + pos["n_closed_shorts"]

pos = pos.with_columns(
    # the level change, as a share of the market
    doi_rel=pos["d_open_interest"] / prev,
    # who is arriving: +1 when every entrant is long, −1 when every entrant is short
    n_new_tilt=(pos["n_new_longs"] - pos["n_new_shorts"]) / arrivals,
    # who is leaving: the mirror image, with the sign flipped so + means "shorts left"
    n_close_tilt=(pos["n_closed_shorts"] - pos["n_closed_longs"]) / exits,
    # how much of the holder base turned over, direction-free
    churn=(arrivals + exits) / pos["n_unique_users"],
)
```

Then the rank IC, with the same guards the study uses: a minimum cross-section width of 25 markets, because a rank correlation over twelve names is not a measurement.

```
def rank_ic(panel, field, forward_return):
    """Cross-sectional Spearman IC per hour; a cross-section needs breadth."""
    return (
        panel.filter(pl.col(field).is_not_null() & pl.col(forward_return).is_not_null())
        .group_by("hour")
        .agg(
            n=pl.len(),
            ic=pl.corr(pl.col(field).rank("average"), pl.col(forward_return).rank("average")),
        )
        .filter(pl.col("n") >= 25)
        .drop_nulls("ic")
    )
```

Do not shrink the coin set within a panel to save on price data. The rank IC needs breadth, and the crypto tail is where the counts are sparse by construction. The axis to cut is time, or drop one panel's price entirely, but not the width.

```
# One closed month, every listed coin, free key:
uv run python oi-composition/oi_composition.py --months 2026-10:2026-10

# The thirteen months behind the tables: same script, wider span.
uv run python oi-composition/oi_composition.py --months 2025-10:2026-10
```

The script reads coins in batches: every coin in a batch is one presigned request, and a single call for a whole month trips the rate limit. If you see a 429, lower the chunk size.

Open interest is a count of contracts, and a contract has two sides. Nothing in the level can tell you which side opened it. That is arithmetic, and no amount of data changes it. The composition can, and it does: who is arriving carries a sign in both of these markets, while who is leaving carries none. Everything past that point belongs to the market structure, not to the ledger. Which is another way of saying: before you compare two markets' open interest, check whether they close at night.

[← Back to blog](https://tesseralytics.dev/blog)
