Above the eight.

Paper 05 found one durable rule: when the past five trading days on NQ have been in the strongest 30 % of the trailing year, double A+ contract size. That rule prints +$28,850 over five out-of-sample years and survives every robustness test we ran.

But the founder asked the obvious follow-up. Is there a way to identify regimes more broadly — extended pullback windows, places to lean short, days where the desk should sit out? Could a single moving average do the work that Paper 05's percentile rule does, but more simply, on more days?

The answer turned out to be yes. The simplest classifier we tested — one EMA, one comparison — beats Paper 05's percentile rule on every measurable axis. Paper 05's days are a 97.6 % subset of the new rule's days. We're retiring the percentile gate on flip day.

We tested every reasonable single-feature regime classifier on daily NQ across the same 7-year window:

• EMAs at 8, 21, 34, 50, 100, 200 bars. For each span: price-above-EMA, slope direction, cross signals, and combined gates.
• Pre-market direction bucketed into five magnitudes (strong-up to strong-down) — the founder's hypothesis that pre-market down moves reverse harder than pre-market up moves.
• Pullback families — day-after-strong-rally, 5-day-up with 1-day-down (consolidation), low-vol-in-trend (ATR(5) < ATR(20)), premarket-pullback gates.
• Extended pullback regime — 1-day return negative AND price below EMA-21 AND realized vol elevated.
• Composite labels across {trend-up, drift-up, range, pullback-in-up, extended-pull, trend-down} built by stacking the above.

For each classifier we ran the full Paper 05 protocol: walk-forward per-year out-of-sample, 500 per-year-stratified random controls with the same N, multiple-comparison correction (Bonferroni and Benjamini-Hochberg), and an economic test against V_FINAL's actual MNQ trade log over five OOS years. A classifier had to clear random-controls at the 95th percentile to count as real edge.

A single feature dominates. If yesterday's NQ daily close is above its 8-period exponential moving average, today's long VWAP-reclaim setups hold 30+ minutes 79.1 % of the time vs 69.7 % on all other days. No fitting, no thresholds, no parameters — just one comparison evaluated at the open. The lift is 8 of 8 years stable on holds rate, 7 of 8 years stable on dollar P&L (the one miss is 2026, a 64-day partial year that is universally negative for V_FINAL).

We then ran the same A+ × 2× upsize protocol Paper 05 shipped. The rule lifts V_FINAL P&L by +$51,328 over the 7-year window — almost exactly 2.4× Paper 05's lift on the same framing. Random controls clear at the 99.0th percentile at 2× sizing and the 99.6th percentile at 3× sizing. Paper 05's `ret_5d top 30%` rule cleared random at 78.2 %.

Paper 05's rule fires on 422 days. The new rule fires on 899. Of those, 412 days are in both — Paper 05's population is a 97.6 % strict subset of the EMA-8 population. The other 487 days are above EMA-8 but not in Paper 05's top 30 % of trailing-year returns.

Those 487 days are the steady grinders. 5-day return is unremarkable — sometimes slightly negative — but daily NQ is making higher closes above its short EMA. Paper 05's percentile rule misses them entirely. They print at 80.1 % holds-30 and $63.99/day, with a random-control percentile of 96.4 % on their own. That's +$31,162 of orthogonal lift Paper 05 never saw.

The mechanical reason is simple. Paper 05's ret_5d filter only activates after an explosive run. EMA-8 activates the moment price gets above it and stays positive through the entire grind — including the long stretches where the trend is real but the trailing-year comp is too noisy to land in the top 30 %.

The first rule of regime overlays is they look great in-sample. Anything that fires on 50 % of days at 2× sizing will scale total P&L by ~1.5× even if the days it picks are random. So the only test that matters is comparing the rule to random-day upsize gates with the same hit count.

We ran 500 random per-year-stratified upsize trials at 2× and 3×. The above_ema_8 rule beat the random-day distribution at the 99.0th percentile at 2× and 99.6th percentile at 3×. The 95th-percentile random benchmark at 2× was +$49,262 — the rule prints +$51,328, comfortably above. The 99th-percentile random benchmark was +$52,576 — the rule lands a hair under that bar at 2× and well above at 3×.

Three sensitivity tests. The rule has to keep working when we wiggle the dials, otherwise it's a curve fit.

EMA span sensitivity is the most important check. If the rule only worked at exactly 8 bars and broke at 7 or 13, it's a coincidence. We tested EMA-5, 8, 13, 21, 34, and 50 on the same protocol. All six landed within ±$3,000 of EMA-8. The rule is span-insensitive — short-to-mid daily EMAs are all telling the same story. The 8-bar choice is convention, not magic.

The founder's original question included pre-market direction. If pre-market goes down before the open, how often does it reverse during RTH? We tested it.

The asymmetry is real. Strong-down premarket sessions V-reverse 60 % of the time vs strong-up at 49 % — uncorrected z = 2.49, p = 0.013. After Bonferroni across the 30 signals we tested alongside it, the finding is borderline. Random-control at 2× upsize is 61 % — well below the 95 % bar we require to ship. Two of eight years invert (2021 and 2025), which kills strategy use.

What we do with it: a footnote on the discretionary short-side checklist from Paper 05. Strong-down premarket sessions V-reverse roughly three out of five times — treat the open as a possible long setup if the structural rails (above EMA-8, recent highs intact) align. Without those rails, default to bear bias and respect the 12:30 cutoff. Discipline rule, not strategy rule.

Five regime ideas tested clean on at least one metric and were dropped on the economic check.

• Q1 strong-down → reverse-long rule. Real asymmetry, but random-control at 2× is 61 % — below the 95 % bar. Lives as a footnote on the human checklist.
• Q3 H1: day-after-+1.5%-rally upsize. 95.5 % holds-30 — the cleanest single-day number in the entire study. But it only fires ~25 days/year, random-control at 2× is 65 %, and most of the days are already inside the EMA-8 above set. Reserved as a possible 3× super-upsize layer once EMA-8 has 30+ live trades.
• Q3 H3: 5-day-up with 1-day-down (consolidation). Holds-30 was worse than baseline. The setup that looks like consolidation in an uptrend is more often distribution. Don't use the naive definition.
• Q4 extended-pullback skip-longs filter. The regime is decisively bearish (45.7 % holds vs 69.7 % baseline, mean RTH move −96 MNQ pts). But skip-filtering longs on these days loses ~$7,400 over seven years — V_FINAL's bias filter has already absorbed the signal. Same selection effect Paper 05 documented.
• Trend-down regime as a skip filter. The composite trend_down regime is the only label with negative average P&L (−$9/day). Skip-filtering it loses ~$1K over seven years on net, only 4/8 years positive. Discipline gate eats the signal again.

The pattern from Paper 05 holds in this paper too: filters that try to suppress bad days lose money because the bias filter already priced them in. Filters that amplify good days print.

Paper 05 ships a rule. This paper ships a strict generalization of it. The graceful path is a flip:

• On flip day: the regime gate becomes above_ema_8. Computed once per day at 9:30 ET from NQ daily continuous data. If the prior daily close is at or above the lagged 8-period EMA, the day's A+ entries fire at 2× contract size. Otherwise baseline.
• What goes away: the percentile lookup against the trailing 252-day distribution. Replaced by one EMA comparison.
• What stays: every other V_FINAL rail — circuit breakers, 15-minute bias filter, 12:30 cutoff, T2 sizing. None of those touch the regime gate.
• The transition is monotonic. Every day Paper 05 flagged is also flagged by EMA-8 (97.6 % overlap; the 2.4 % difference is at-EMA edge cases). No day loses sizing in the flip; 487 additional days gain it.

For the desk and for any subscriber reading along: this is the one-pager. The discipline lives here.

AT 9:30 ET — read NQ daily continuous:

  daily_close[D-1]   ≥   EMA(daily_close, 8)[D-1]
  ──────────────────────────────────────────────
  TRUE   →  REGIME-UP. A+ entries fire at 2× contract size.
  FALSE  →  BASELINE. A+ entries fire at base size.

NOTHING ELSE CHANGES:
  ▸  T2 (smaller confluence trade) keeps base size always
  ▸  15-minute EMA-21 bias filter runs as is
  ▸  12:30 ET cutoff applies
  ▸  Circuit breakers untouched

REGIME-UP CONTEXT (footnotes for the human side):
  ▸  Strong-down pre-market (≤ -0.5%) reverses ~60% of
     the time — possible long opportunity if structural
     rails hold. Default to bear bias if they don't.
  ▸  Day after a +1.5% NQ rally (rare): 95.5% holds.
     Don't upsize beyond 2× yet — reserve until EMA-8
     has 30+ live fires.

NOT REGIME-UP CONTEXT:
  ▸  Don't skip longs. The bias filter already cuts
     most of them. The ones that fire pay 55%+ at +$84/trade.
  ▸  Use Paper 05's archetype playbook for discretionary
     short setups (bear flag, VWAP rejection, lower high,
     trend-day continuation, failed breakout).
  ▸  Flat by 12:30 ET. No exceptions.

WHAT THE RULE WILL FEEL LIKE LIVE:
  ▸  Fires ~50% of trading days
  ▸  ~8 days/year better holds in every year tested
  ▸  Worst single-year drawdown on the upsize curve: -$5,570
  ▸  The 2026 partial year is universally negative for
     V_FINAL — not a regime-rule failure, a base-rate fact

Three follow-ups this paper deliberately leaves on the table.

• Tiered upsize. EMA-8 (2×) plus Q3 H1 (3× super-upsize on day after +1.5 % rally). Math works in backtest; we don't ship layered structure on a brand-new gate. Revisit after EMA-8 has 30+ live fires.
• Trend-down + market internals. Trend-down is the only regime where V_FINAL is net-negative. Can a second filter — breadth divergence, VIX spike, sector internals — clean it up enough to skip-filter at 95 % random-control? The single-feature tests said no. A two-feature gate might.
• Discretionary short module on extended-pullback regime. Q4 is decisively bearish (197 days, 45.7 % holds, −96 pts RTH mean). Stack Paper 05's five archetypes underneath it. The next paper writes itself, but it's gated on running an automated short backtest first — F33 in the master findings file says programmatic shorts print at 0.4–0.7 RR without a regime gate. With one, the math might change.

The simplest possible regime classifier — is daily NQ above its short EMA? — wins every measurable test we ran. It is the most robust, walk-forward stable, slippage-tolerant, span-insensitive, and absolute-dollar-largest overlay we have found in the 7-year window.

Paper 05 found the right shape. This paper found the cleaner version of it. Paper 05's 422 days are inside this paper's 899 days; the 487 additional days print at 96th-percentile random-control on their own. We retire the percentile gate and ship the EMA on flip day.

One line of code. One daily computation. No fitting. The private-desk regime gate just got smaller and stronger at the same time.