Finding, not used

Leverage crowding

Assets whose leveraged longs are most crowded do worse over the following days — real information that no simple portfolio manages to collect.

The question: Can you measure when too many people are leaning on the same side, and is it good for anything?

What was found

It's the signal that best illustrates the difference between "there is information" and "there is money". The mechanism is prior and concrete: somebody pays the carry, and crowded leverage is what forced liquidations feed on. Measured as a cross-sectional percentile — how crowded this asset's longs are against the rest of the market — the signal has the predicted sign at all three horizons (1, 3 and 7 days), with block-permutation significance, the same sign out of sample, independence from the known factors (momentum, low volatility, size, illiquidity) and a parameter plateau of 16 cells out of 16, negative in AND out of sample. That's the opposite of every earlier supposed finding in this project, where the result depended on the exact cell.

return horizonh = 1h = 3h = 7h = 14window1-0.049-0.064-0.046-0.0583-0.048-0.061-0.046-0.0657-0.039-0.049-0.043-0.06614-0.034-0.044-0.045-0.076
The parameter plateau in exploration: four trailing windows × four horizons. A fluke lights up one or two cells; a real effect fills the grid with one sign.analysis/scripts/phase9b_crowding_construction_2026-09-07.py
return horizonh = 1h = 3h = 7h = 14window1-0.009-0.019-0.031-0.0433-0.016-0.028-0.043-0.0507-0.019-0.028-0.046-0.05714-0.011-0.027-0.045-0.060
The same grid in the holdout, which was never used to choose anything. 16 of 16 keep their sign — weaker, but not one switches sides.analysis/scripts/phase9b_crowding_construction_2026-09-07.py
-0.069-0.050-0.031-0.0120.007-0.049-0.023-0.0141d-0.061-0.033-0.0253d-0.046-0.037-0.0297d13 symbols (exploration)56 symbols (exploration)56 symbols (holdout)
What happens going from 13 to 56 symbols: the IC shrinks. That's what a larger sample does to an optimistic estimate, and it's good news about the method rather than bad news about the signal.analysis/scripts/phase9_crowding_wide_2026-09-07.py

Try it yourself

The information coefficient (IC) is the rank correlation between a signal and future return. Here you can see how small an IC of 0.03 is once translated into the only thing that matters: how much relative return it implies.

Expected relative return % ≈ IC × signal z × Return dispersion
Expected 7-day relative return

It's an AVERAGE TENDENCY over hundreds of observations, not a prediction about one asset on one day. At the measured IC (−0.037) and an asset at the crowding extreme, the expected edge is a tenth of a percentage point: below what the fee to get in and out takes.

Illustrative example numbers for practice — not real data.

How it was tested, step by step

  1. The hourly premium index is downloaded from Binance's public archive — the basis between the perpetual and spot — for dozens of pairs. The funding rate is a coarse version of this same measure: it settles every 8 hours and is capped at ±0.75%.
  2. Every day all assets are ranked by their basis level, and each one's signal is its relative POSITION, not its absolute value. That removes at the root the problem of having to beat an asset with positive drift: you only have to rank assets against each other.
  3. It's checked against being a known factor in disguise: orthogonalized against momentum, low volatility, size and illiquidity all at once, it remains significant at 1 and 3 days.
  4. And the panel is widened from 13 to 56 symbols with the same pre-registered hypothesis and the same parameters. The IC shrinks — as expected, the original 13 were the ones with the longest history and deepest liquidity — but it's confirmed: 16 of 16 cells are still negative in both windows.
What this does NOT say. And here's the uncomfortable verdict: the portfolio falls apart. Rank-weighted and dollar-neutral, it goes from a Sharpe of 0.70-1.49 in exploration to between −0.06 and −0.60 in the holdout, with a NEGATIVE gross break-even cost out of sample. TWO standard constructions were declared before testing them — volatility scaling and restriction to the 25 most liquid — plus their combinations: 12 of 12 turn negative in the holdout. A third won't be tried, because trying constructions until one works is exactly the mistake this project documented in itself. The Sharpe of 0.27 the 13-symbol panel gave was a small-book artifact: widening the sample did precisely its job, shrinking an optimistic estimate instead of flattering it. Add that 2025 was a year with the sign flipped.

Tracking

This finding has no live number: it was measured on archived data, and putting a chart here would imply continuous tracking that doesn't exist. The figures are above, with the script that produced them.

← All research