Is that the high? We measured it. Then we killed it.
A calibrated, walk-forward-validated curve for whether the current extreme holds — 1,403 NQ sessions, 100,990 observations. Then the structureless null: 97% of the curve is geometry. The kill is worth more than the curve.
Every trader asks it daily, usually around mid-morning: is that the high?
We measured it across 1,403 NQ sessions. Then we ran the test that most published stats never face — and most of the answer died. This drop is both halves, because the kill is worth more than the curve.
The curve
P(the extreme currently in place ends up as the day’s extreme), by how long it has stood unbeaten. Tick-derived 1-minute NQ, RTH sessions 2020-11-27 → 2026-07-27, sampled every 10 minutes from 10:00 on, highs and lows pooled — 100,990 observations.
| Unbeaten for | P(it’s the day’s extreme) |
|---|---|
| under 15 min | 16% |
| 15–30 | 37% |
| 30–45 | 47% |
| 45–60 | 55% |
| 60–90 | 61% |
| 90–120 | 68% |
| 2–3h | 75% |
| 3h+ | 89% |
The unconditional base rate is 61.7% — what you’d say knowing nothing but that an extreme exists. The curve swings from 16 to 89 around it.
And it validated. Walk-forward, every session scored only on data that came before it, 86,614 observations: largest calibration miss in any bucket, 1.6 percentage points. Brier skill +26.9% over always quoting the base rate. Stable across three eras, strongest in the most recent. By every standard test, this was a keeper.
The kill
One question remained: what would a market with no structure at all do?
The running maximum of any random path gets harder to break as it ages. That requires no market, no auction, no participants — it’s a property of noise. So the honest opponent isn’t “always quote 61.7%.” It’s a path with NQ’s exact volatility and zero structure. We built two: each real session’s own 1-minute returns shuffled within the session, and a volatility-matched random walk.
| Unbeaten for | Real NQ | Shuffled returns | Real − shuffled |
|---|---|---|---|
| under 15 min | 16.2% | 15.0% | +1.2 |
| 15–30 | 36.9% | 32.1% | +4.7 |
| 30–45 | 46.8% | 39.8% | +6.9 |
| 45–60 | 54.3% | 47.0% | +7.3 |
| 1–1.5h | 60.9% | 53.9% | +6.9 |
| 1.5–2h | 68.1% | 61.4% | +6.7 |
| 2–3h | 75.8% | 70.0% | +5.7 |
| 3h+ | 88.9% | 85.6% | +3.3 |
The curve spans 72.7 points on real data — and 70.6 points on shuffled noise. About 97% of the curve is mechanical. Scored properly:
- Skill vs. a constant base rate: +27.6% ← what we originally reported
- Skill vs. the structureless null: +1.5% ← the honest number
About 97% of the curve is mechanical. Skill vs. the structureless null: +1.5%.
What survives
A small, consistent residual: through the middle buckets, real NQ extremes hold 3–7 points more often than pure noise implies — plausibly trend persistence. It’s real. It’s worth +1.5% forecast skill, which is close to nothing.
So the honest answer to “is that the high?” is: the age of the high tells you almost exactly what a random walk would tell you. A 10-minute-old extreme is noise and a 3-hour-old one is nearly settled — but that’s a property of elapsed time under volatility, not market information. Useful as a yardstick for your own expectations; not an edge, and we won’t dress it as one.
The method lesson (this is the actual product)
A weak baseline manufactures skill. “Beats a constant” is nearly always available and nearly always meaningless. The curve passed walk-forward calibration, era-stability, and a session-blocked bootstrap — and still turned out to be 97% geometry. Every conditional probability should be scored against a structure-destroying null — shuffled returns or a matched random walk — before anyone quotes it as an edge.
Ask that question of the next stat someone sells you.
— Jordan Dallas, AXIS Desk
Research/education, not advice. Futures trading involves substantial risk of loss.