Spot price tells you what one asset traded at. The derivatives market tells you what it cost to be positioned, how much borrowed money is riding on the outcome, and how crowded one side of the boat has become. That second set of facts is published continuously, for free, by every major venue - and it is where machine learning has the most defensible job in crypto, provided you are honest about what the job is.
It is not calling direction. It is measuring fragility. Those sound similar and they are completely different products.
Three numbers that describe positioning
Perpetual futures have no expiry, so something has to tether them to spot. That something is the funding rate: a periodic payment from longs to shorts, or the reverse, sized by how far the perp has drifted from the index. When funding is positive, leveraged longs are paying to keep their position open. When it is negative, shorts are paying. It is the clearest published price of crowd conviction in any market we know of, and it updates several times a day.
Open interest is the total notional sitting in unclosed contracts - how much leverage exists, not how much has traded. Read together with price it separates two very different moves. Price up with open interest rising means new money opening longs. Price up with open interest falling means shorts closing out, a rally made of exits rather than entries. Same candle, entirely different structure underneath.
Basis is the third: the gap between futures and spot, annualised. It is the carry the market is paying for term exposure, and its shape across expiries says whether demand for leverage is a front-month spike or a settled condition.
None of these is a secret. All of them are structural facts about who is positioned and what it costs them, which is exactly the sort of thing a model can use and a headline cannot.
What the data actually predicts, and what it does not
Here is the finding that matters, and it disappoints most people who arrive at this data hoping for signals.
Funding and open interest carry real information about the size of coming moves and very little reliable information about their direction. High positive funding plus concentrated open interest describes a market where a modest adverse move triggers liquidations, which force market orders, which move price further, which trigger more liquidations. That is a fragility condition. It says a large move is more likely than the recent calm suggests. It does not say which way, and the honest version of the model stops there.
The contrarian shortcut - funding is very positive, therefore sell - is the trade everyone discovers in their first month. It works often enough to feel like a rule and fails at exactly the moments that matter, because a crowded trade can stay crowded through an entire trend and the funding cost is priced by people who can see the same number you can. What survives out of sample is the humbler statement: leverage build-up widens the distribution of outcomes. That is a volatility forecast, and it is genuinely useful for sizing, for stop placement, and for knowing when not to be casual.
The data traps, which are unusually nasty here
Derivatives data looks tidy and is full of edges that will quietly wreck a study.
Funding intervals differ by venue. One exchange settles every eight hours, another every hour, and some publish an annualised figure while others publish the per-interval rate. Averaging across venues without normalising the interval produces a composite that means nothing, and it will look perfectly reasonable on a chart.
Funding is clamped. Exchanges cap the rate, so in genuinely extreme conditions the published number stops rising while the real imbalance keeps growing. Precisely when the signal matters most, it saturates. A model trained on clamped data learns that stress has a ceiling. It does not.
Open interest is denominated inconsistently. Coin-margined contracts report open interest in coin terms, so a falling price mechanically shrinks the dollar figure with no position actually closing. Mixing coin-margined and stablecoin-margined venues into one dollar series manufactures a deleveraging event that never happened.
Exchange-reported open interest includes internal market-maker inventory that is hedged elsewhere, so some of what reads as directional risk is nothing of the kind. And dated futures roll: naive concatenation across contract months creates a phantom collapse in open interest every quarter.
We have hit versions of all of these. None of them announce themselves. Each one produces a clean-looking series with a defect buried in the middle, which is the most expensive kind of data problem there is.
The history is shorter than it looks
Perpetual futures data in real depth covers a few years, and within that span the market structure changed repeatedly: new venues, changed margin regimes, an entire cohort of leverage that appeared and then vanished across the 2022 failures. The number of genuinely independent regimes in the sample is small, even though the row count runs to millions.
This is the same discipline we apply to forecast bands. The count of observations is not the count of evidence. Overlapping windows and correlated days inflate the first while leaving the second alone. When a derivatives study reports a strong result, the question is how many independent market regimes stand behind it - and the answer is usually a single-digit number that nobody printed.
Where this feeds our forecasts
Positioning data does exactly one job in our pipeline, and it does it in the width of the band rather than the placement of its centre. When leverage is concentrated and funding is stretched, the empirical distribution of forward moves we draw from is a wider one, and the published range reflects that. When positioning is calm, the band tightens. The centre stays where the measured distribution puts it. We do not tilt it with a story about the crowd being wrong.
And it gets committed before the fact. The forecast is serialised, hashed with SHA-256, anchored to the Bitcoin blockchain through OpenTimestamps, and published only afterwards, then scored openly on coverage with the misses on the same page as the hits. Positioning data is unusually seductive - it feels like seeing the other players' cards - which is exactly why the claims built on it need to be locked before the outcome rather than narrated after it.
The deflation
Derivatives data will not tell you where price is going. It will tell you how much of the market is levered into a view, what that view costs to hold, and how violently the exits would work if it broke. Read that way it is one of the better inputs in crypto. Read as a direction signal it is a well-documented way to be right repeatedly and then wrong once, expensively.
No method reliably beats a liquid market, and anyone promising that is selling something. What honest positioning analysis buys is not foresight. It is a band whose width means what it says on the days that turn out to matter.
Educational content - not financial advice, and not a betting tip.