What a perpetual future actually costs

A perpetual future never expires, so nothing forces it to converge on spot. Instead it pays funding — an hourly cash transfer between longs and shorts. These two charts are the price of bitcoin, and the price of holding bitcoin synthetically.

loading…

drag to pan · scroll or pinch to zoom ·
BTC-USD spot Coinbase, from 2015-07-20 close up close down
Hyperliquid BTC perp — premium and funding hourly, from 2023-05-12 premium (perp − spot) funding paid clamp corridor ±5bp

Both series are in basis points per hour — one axis, one unit. Positive premium means the perp trades above spot and longs are crowded.

Read the second chart

The grey corridor is the part almost nobody draws, and it explains the shape of everything else. Hyperliquid sets funding as

funding = premium + clamp(baseline − premium, ±5bp)

Work the algebra. Whenever the premium sits within ±5bp of the baseline, the two premium terms cancel exactly and funding equals the baseline — of all hours on record. Inside that corridor, funding is completely blind to the premium. There is no restoring force. The perp can drift away from spot and nothing pulls it back, because unlike a dated future there is no expiry to force convergence.

So the orange line is flat as a table until the blue line escapes the grey band — and then it moves. That is the entire mechanism, and it is why the premium wanders as freely as it does.

Does the premium predict price?

No. It is worth being blunt, because funding rates are sold as a signal constantly. Correlation of premium against log return, over aligned hours:

windowtrailing returnforward return

Premium tracks the integral of past returns over several days — positioning that accumulated while price ground in one direction. It carries essentially nothing about the future. It tells you where the crowd is, not where price is going.

The regime, not the constant

Funding is often described as structurally positive, and over the full record it is: longs paid in of hours, for a year. But the average hides a total inversion of character.

yearhoursmean premiumfunding paidannualisedpremium > 0

2024 was a crowded-long market — the perp traded above spot nearly three quarters of the time and longs paid around a quarter of notional a year for the privilege. 2026 is its mirror: the perp sits below spot almost always, and carry has collapsed. Same instrument, same formula, opposite world.

Table view — the most recent 48 hours
hour (UTC)spot closepremium (bp)funding (bp/hr)annualised

Where the numbers come from

Two public endpoints, no API key, no account. Both send access-control-allow-origin: *, so this page tops the static history up with a live tail straight from the browser.

seriessourceendpointcap per request
BTC-USD spot candlesCoinbase Exchange/products/BTC-USD/candles300 candles
perp premium + fundingHyperliquid/info · fundingHistory500 hours

Both caps are why scripts/backfill.mjs is mostly pagination. It runs incrementally — a scheduled workflow refetches only the tail and merges. scripts/analyse.mjs derives every statistic quoted above into data/stats.json, so the prose cannot drift from the data.

Honest limits

Spot is Coinbase's BTC-USD; Hyperliquid's own oracle is a composite of several venues, so the premium here is not exactly markPx/oraclePx − 1 at the same instant. The published funding is computed from a time-weighted average of premium samples taken through the hour, while the premium field is a single figure — reproducing funding from it gives a median error of but a tail out to . That gap has a visible signature, and it is the best evidence the corridor story is right: the identity holds of the time when the premium sits deep inside the corridor and only at its edge, because near the boundary the hidden average crosses where the published figure did not. The clamp and baseline are Hyperliquid's parameters, not universal: other venues use different widths and 8-hour intervals. And none of this is advice — it is a chart of what a thing has cost.