The first robust single-signal stack

The session’s previous “first composite to stack” claim was PR #746’s three_clock_vol_weighted. At N=10 it stacked above both parents and baseline. At N=50, it collapsed to ties. PR #753 documented the collapse; the session-summary updates treated that finding as a small-N noise example.

PR #815’s mom_ma_composite (baseline raw return + MA-crossover) is different. It stacks at both N=50 and N=100 on single-signal data.

The result

N=50 single-signal:
  baseline:         +0.767
  ma_crossover:     +0.636  (parent 2; weaker)
  mom_ma_composite: +0.881  ← stacks (+0.114 vs baseline)

N=100 single-signal:
  baseline:         +0.863
  ma_crossover:     +0.742
  mom_ma_composite: +0.934  ← stacks (+0.071 vs baseline)

The stack persists across N=50 → N=100. Δ shrinks slightly (+0.114 → +0.071, 1/√2 convergence as expected), but remains robustly above zero.

Why this stacks where others didn’t

Three composites that DIDN’T stack on single-signal at N=100:

1. three_clock_vol_weighted: parents correlated at +0.71 — same band as mom_ma_composite. Collapsed to ties at N=50.

2. three_clock_portfolio_vol: parents correlated at +1.00 (gate dormant). Interfered.

3. vol_weighted_mean_revert: parents anti-correlated (-0.42). Hedged, didn’t stack.

What makes mom_ma_composite different?

The two parents (baseline log(P/P_lookback) and ma_crossover (MA_short - MA_long) / MA_long) measure THE SAME underlying signal at TWO DIFFERENT smoothing scales. The composition isn’t combining INDEPENDENT alpha — it’s averaging two different windows on momentum.

• baseline: instant 20-day return.
• ma_crossover: smoothed 5-vs-20-day average difference.

When prices have moved up monotonically, both signals agree. When prices have moved up with noise, baseline reads positive but ma_crossover reads small. Their average rejects the noise-dominated case, picking up only the strongest signals.

This is noise reduction via redundant measurement — the classic stacking pattern when both parents look at the same underlying quantity from different angles.

The contrast with three_clock_vol_weighted

three_clock_vol_weighted had similar pairwise correlation (+0.71) and similar shape (score-stage composite). But its parents (three_clock_momentum and vol_weighted) measure DIFFERENT things:

• three_clock_momentum: composite-horizon momentum (mom_5 + mom_20 + mom_60).
• vol_weighted: per-symbol vol-normalised momentum.

These two parents see different aspects of the data: the first reads horizon-decomposition; the second reads risk-adjustment. Both useful, but the combination is multiplicative when one parent’s signal cancels the other’s.

mom_ma_composite’s parents are both ABOUT THE SAME THING (20-day momentum) measured DIFFERENTLY. That’s the redundant- measurement stack.

Refining the pairwise rule (towards a 6th iteration?)

ADR-0062’s 5-iteration rule says:

Composite stacks iff:

1. Low POSITIVE correlation (0 to +0.5).
2. Same-stage.
3. Mechanism preserves both decisions.

But mom_ma_composite is at +0.70 (above the “0 to +0.5” threshold). The rule’s “marginal band” (0.5-0.8) was supposed to predict “stack or flat” without commitment.

This finding suggests refining the marginal band’s prediction:

In the +0.5 to +0.8 band, the composite stacks IF the parents are measuring the same underlying quantity at different granularities (redundant measurement). It is flat if the parents measure different signals at the same granularity (multiplicative interference).

Whether this distinction is reliable enough to ship as the 6th iteration depends on future composite tests. The current evidence: 1 stack (mom_ma_composite) vs 1 tie (three_clock_vol_weighted) in the +0.5-0.8 band.

What this rules out

• Not “all +0.7 composites stack.” three_clock_vol_weighted was +0.71 and tied. The “marginal” band is mixed.

• Not “mom_ma_composite is the harness leader.” ts_momentum still leads at +0.995, +0.061 above the composite. The composite stacks ABOVE baseline but BELOW the entry-rule arm.

• Not “real markets work this way.” Synthetic-specific. On real data, baseline’s 20-day return and ma_crossover’s MA difference may correlate differently and the stack may vanish.

What to ship next

Three follow-ups:

1. Regression test pinning the stack. mom_ma_composite > baseline at N=30 by >0.03 Sharpe. Calibrated similarly to PR #782’s ts_momentum test. (Already an obvious next PR.)

2. Test the refined marginal-band rule. Find a pair of arms in the +0.5-0.8 correlation band that measure DIFFERENT things at the SAME granularity. Predict: flat. Build + measure.

3. Real data. Synthetic baseline + ma_crossover stack; does real-market mom + MA-crossover stack too?

The closing observation

The session’s first ROBUST single-signal stack arrived at PR #815, 100+ PRs after the harness’s birth. The pattern was right there all along: combine two different smoothings of the same momentum signal. Five iterations of the pairwise rule brought us close — the 5-iteration ADR-0062 predicted “marginal,” and the finding lands on the stack side of marginal.

The platform has matured to the point where it surfaces robust findings. The next operator inherits this — and the ability to run the same loop on real data.