Null result

The other null families

Cross-sectional momentum, seasonality, macro, stablecoins, ETF flows, futures positioning and options skew — tested with the corrected method.

The question: Outside charts and on-chain indicators, is there anything on the other axes?

What was found

These families are grouped here because they share a verdict, not because they're the same thing. The most interesting is seasonality: the effects are LARGE and universal in exploration — 21:00 UTC gives +6.12 basis points an hour with agreement across 14 of 14 assets, and Saturday +1.94 — and they're still not tradeable. The hourly version rotates 3,650 times a year, giving a 365% annual cost, and 0 of 14 symbols beat their baseline. The day-of-week version is cheap but its edge is +0.030 of Sharpe: real and irrelevant. And 03:00 UTC, one of the strongest effects in exploration, FLIPS SIGN out of sample.

-7.82-3.96-0.11+3.75+7.61+6.12+3.1521:00 UTC-6.33+2.2603:00 UTC-5.64-0.3914:00 UTC+1.94+0.95SaturdayExplorationHoldout 2023+
Hourly and weekly seasonality: basis points per hour in exploration against the holdout. Large effects, universal across 14 of 14 assets... and one of them flips sign out of sample.analysis/scripts/phase5_intraday_structure_2026-09-07.py

Try it yourself

Why a real effect can be worth nothing: drag the edge per rotation, the cost and the frequency. Hourly seasonality rotates 3,650 times a year.

Annual result % = (Edge − Cost) x Rotations / 100
Net annual result

What is paid in fees per year:

With the default values, the strongest hourly effect measured — a real effect, with agreement across 14 of 14 assets — pays 73% a year in fees. Dropping the frequency to one rotation a week makes it cheap and, at the same time, irrelevant.

Illustrative example numbers for practice — not real data.

How it was tested, step by step

  1. Cross-sectional momentum and reversal over a 65-symbol panel: no significant information coefficient (all with p above 0.51). It's also the most replicated anomaly across every asset class, so the null here says something about the panel — it's only wide from 2020 — as much as about crypto.
  2. Six cross-sectional factors with a mechanism (low volatility, beta, Amihud illiquidity, size, volume shock, distance from the 52-week high): no significant IC under block permutation, despite Newey-West t-statistics of up to 3.75. That discrepancy is itself a methodological finding.
  3. Stablecoin liquidity (30-day growth: IC 0.031, p=0.80; inverted SSR: IC −0.066, p=0.69), cross-asset macro with DXY, gold and the S&P, the volatility risk premium as a direction signal, and CFTC futures positioning: all null.
  4. ETF flows: the only one with a nuance. It was the first positive from a never-tested source, but re-measured with the corrected method and an internal split (the global holdout is useless because the ETF starts in 2024) it gives IC 0.095 at 30 days with p=0.15. Better than the baseline and underpowered: 6 quarters.
What this does NOT say. There's a partial exception that isn't closed: top-trader positioning. The long/short ratio of the largest accounts gives the correct sign at all three horizons with the same sign out of sample, and its correlation with the already-validated crowding signal is only 0.275 — that is, it measures the same mechanism by another route, which is what makes it count as confirmation. But under strict multiplicity correction none of the run's 12 cells survives at q=0.05 or q=0.10 (at q=0.20 exactly the ratio's three do), and the alt history only reaches December 2021. The nuance it does contribute is informative: the AMOUNT of leverage predicts nothing (open-interest growth, new-long intensity: null, and sign-flipped out of sample); what predicts is its COMPOSITION.

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