In use

The basis carry

Collecting the interest leveraged longs pay, with no exposure to price direction at all.

The question: Is there any income in this market that doesn't depend on getting anything right?

What was found

After 6.68 million hypotheses trying to predict price, the idea that worked was to stop predicting. A perpetual future never expires, so to keep it from drifting away from spot the exchange makes longs pay shorts every eight hours whenever there are more longs than shorts. Someone who holds real bitcoin and is simultaneously short the perpetual neither gains nor loses from price — what rises on one leg falls on the other — but collects that payment. It isn't a forecast: it's a contractual cash flow. Measured over 2,417 days with both legs modeled and 10 basis points of fees, the Sharpe is above 10 in EVERY single year since 2020.

0.0%4.2%8.3%12.5%16.7%7.8%0.9%202014.9%1.4%20211.3%3.0%20223.1%4.0%20235.2%4.2%20241.8%4.3%20250.5%4.4%2026Net carry on capitalUS 10-year Treasury
Net carry on capital deployed, year by year, against what idle money earned. Since 2022 the bond wins almost every time.analysis/scripts/phase15b_carry_reality_2026-09-07.py

Try it yourself

The deciding question: is anything left after subtracting what idle money earns and the fees? Drag the three pieces.

Net excess = Carry ÷ 2 (capital, not notional) − Risk-free rate − Fees
Excess over leaving the money still

Carry on capital, before comparing:

The ÷2 isn't an arbitrary conservative haircut: a real position needs the spot leg plus margin for the short, so the capital tied up is roughly twice the quoted notional.

Illustrative example numbers for practice — not real data.

How it was tested, step by step

  1. Bitcoin is bought spot and the same quantity is simultaneously sold in the perpetual future. The net bitcoin position is zero: what price does is irrelevant.
  2. Every eight hours funding is collected (or paid). BOTH legs are modeled: the funding itself and the spot-perpetual convergence, using the real hourly premium index, not an approximation.
  3. It's measured on CAPITAL DEPLOYED, not on notional. That's the difference that usually inflates this number by two wherever it gets published.
  4. And it's compared against what that same capital earns sitting still: the US 10-year Treasury. Income only matters against its alternative.
WindowNet return/yearVolatilitySharpeWorst day
Exploration (<2023)8.76%0.80%11.00−0.63%
Holdout 2023+3.64%0.24%15.21−0.06%
Full history5.93%0.58%10.28−0.63%

A Sharpe of 15 doesn't mean it's a good idea: it means it's very stable. What decides is the level, and the level is below the bond.

What this does NOT say. The honest conclusion is both things at once: it's the only genuine, repeatable income the project has found, and today it isn't worth doing. Over the last 365 days it returned 0.85% against the bond's 4.29% — an excess of −3.45%. The trade has been competed down below Treasury bills. That's why it's published as a LIVE number rather than as a strategy: the opportunity opens and closes in stretches, and the number says which of the two it is today. And adding the convergence leg changes the result by exactly nothing: the premium reverts to its mean, so its cumulative contribution cancels out.

Tracking

This finding does have a real number that gets published and kept updated.

Data last updated on 2026-09-09.

← All research