One number at 10:30 tells you what kind of day it is

Last week’s drop was a kill. This week is a survivor — a stat that went through the same null that executed our trust curve, and came out alive. Modest, real, and useful by lunch.

The statistic

At 10:30 ET, take one number: how far price has traveled from the RTH open, divided by ATR (mean of the prior 10 completed RTH ranges — no lookahead). “Directional day” is defined up front as |close − open| ≥ 0.6 ATR; 36.6% of sessions qualify — that’s the base rate to beat.

Read the last column before the middle one. That’s the same statistic computed on shuffled versions of each session — same volatility, zero structure. Travel-so-far is arithmetically part of travel-at-the-close, so some “predictability” is guaranteed. The floor column is that arithmetic; only the gap above it is market.

The score

• Skill vs. a constant base rate: +10.2%
• Same pipeline on shuffled paths: +3.2% (the mechanical part)
• Genuine skill: +7.0%

Stable across eras (2022–23: +10.5%, 2024: +6.5%, 2025–26: +10.1%). Session-blocked bootstrap puts the quietest quintile at 16.7% [11.7–22.1] and the busiest at 61.6% [54.8–68.2] — non-overlapping intervals. Walk-forward out-of-sample, 987 sessions, fitted only on prior data. 1,387 sessions total; the parquet was audited against exchange daily files first (median high/low difference across all 23 contracts: 0.00 points).

The asymmetry that matters

The quiet end is the strong end. Under ~0.1 ATR of travel by 10:30, you’re looking at a balance day roughly 85% of the time — and the random-path floor up there is only 74%. The busy end is weaker: 59.7% directional against a 49.6% floor is a 10-point residual, half of it geometry.

So the read isn’t “a fast tape means a trend day” — it’s the inverse:

A dead tape at 10:30 is the most trustworthy character call NQ offers.

What died on the way (published so you don’t rediscover it)

• “The opening extreme still holds” as a trend-day predictor: AUC 0.527 — the weakest thing we tested. It describes trend days in hindsight, predicts nothing.
• “Range so far”: AUC 0.696 — looks nearly equal to travel’s 0.705. Its shuffled floor is 0.666 vs. travel’s 0.593. Almost entirely hollow. Two predictors that look identical on a leaderboard; one is real, one is geometry.
• Adding more features made it worse in genuine terms — extra features inflate the mechanical floor faster than the signal.
• The fitted probability curve over-quotes at every level (worst gap −7.1 points). We ship the empirical quintile lookup instead. Ordering is right; confidence isn’t.

How the desk uses it

It says nothing about direction — it’s a magnitude-and-character read only. A quiet 10:30 argues for balance-day behavior: fade-friendly, both-extremes-still-moving, mean-reverting tape. A busy 10:30 argues mildly for range extension and mostly for humility. Context for what kind of day you’re in, not a reason to enter anything.