The intervention-point rule, confirmed by a fourth experiment

PR #731’s discipline rule:

When N independent shapes of the same intervention all fail in the same direction, the intervention itself is the wrong operation for the data. Move to a different intervention point.

PR #732 shipped XsMomentumWithPortfolioVolGateStrategy to test this. The intervention-point change: instead of per-symbol filtering of the cross-sectional ranking (which all three variants did), gate the AGGREGATE PORTFOLIO size on a universe-wide vol signal.

PR #734 ran it. Here’s the comparison.

The four-variant table

Mode:                  I.I.D.       Clustered (cluster=0.8)
baseline:              +0.961       +1.348
vol_regime_filter:     +0.791       +0.526
vol_transition_filter: +0.540       +0.329
vol_penalty:           +0.591       +0.634
portfolio_vol_gate:    +0.772       +1.148  ← new

The clustered-vol gap to baseline, by intervention shape:

Variant	Gap	What it recovers
vol_regime_filter (per-symbol level)	−0.822	0% (worst)
vol_transition_filter (per-symbol change)	−1.019	(worse than worst)
vol_penalty (per-symbol score)	−0.714	13%
portfolio_vol_gate (cross-symbol)	−0.200	76%

Same alpha source (vol clustering), same data, same momentum factor, same gross_leverage. The only thing that changed was where the vol signal entered the strategy — at the per-symbol filter, at the per-symbol score, or at the portfolio-aggregate scale.

Cross-symbol won by a wide margin.

What the result confirms

The intervention-point rule was the right framing of the problem. The previous shape-iteration (level → transition → penalty) was exploring the wrong axis — those are three shapes of the same intervention point (per-symbol filtering of the cross-sectional ranking). They all failed for the same reason: per-symbol filtering shrinks the universe in a way that costs more Sharpe than the noise-reduction recovers.

The cross-symbol intervention point preserves the full universe for the ranking and intervenes only at the dollar scale. The cross-sectional ranking’s robustness is intact; only the magnitude shrinks.

What “75% recovery” tells us

Not 100%. The cross-symbol gate still costs ~0.2 Sharpe vs baseline on clustered data. This says the cross-symbol vol signal isn’t FREE — sizing down the portfolio during high-vol windows misses some of the upside that the unscaled baseline captures. The trade-off:

• Baseline: full exposure always; benefits from rare positive-vol regimes.
• portfolio_vol_gate: sized-down exposure during vol clusters; misses some upside but reduces drawdown in negative-vol regimes.

On the synthetic, baseline is mildly better on mean. The operator’s choice depends on whether they want the smoother ride (portfolio_vol_gate’s stdev 0.610 vs baseline’s 0.677) at the mean cost.

What this rules out

• Not “cross-symbol always beats per-symbol.” The cross-symbol variant still loses to baseline. It JUST loses by less than the per-symbol variants. The intervention-point shift is partial good news, not full vindication.

• Not “vol intervention works.” Baseline still wins on mean Sharpe across both modes. The intervention-point rule narrowed the gap, but the underlying “vol-aware beats vol-blind on this synthetic” claim is still rejected.

• Not generalisable to real data. Real markets have documented vol-regime alpha; the synthetic doesn’t. The cross-symbol intervention might beat baseline on real data, but the synthetic can’t tell us so.

What to ship next

Three follow-ups, in priority order:

1. Tune the cross-symbol thresholds. universe_vol_threshold=1.2, alpha=1.0, min_scale=0.2 are defaults. A sweep across these might reveal a sweet spot where the cross-symbol variant beats baseline.

2. Compose with momentum scoring. PR #710’s pairwise-correlation note showed composites can stack on independent signals. A XsThreeClockMomentumWithPortfolioVolGateStrategy would test whether the portfolio gate composes with composite-horizon momentum (where the three-clock arm already beats baseline).

3. Real data. The vol-clustering hypothesis is real-market documented. The synthetic results show the SHAPE of the intervention that’s likely to work; verifying it works on real bars is the load-bearing test.

The discipline rule, refined

PR #731’s original rule:

When N independent shapes fail, move to a different intervention point.

PR #734’s confirming evidence:

Moving the intervention point recovered most (but not all) of the lost Sharpe. The remaining gap suggests the intervention still has a cost — just a much smaller cost than the original shapes.

Combined rule:

When N independent shapes fail in the same direction, change intervention point — but expect partial recovery, not full. The shape-axis was exploring within a wrong assumption; the intervention-point change tests the assumption itself, but the remaining gap shows the underlying value-add is bounded.

A corollary: in this case, the “underlying value-add is bounded” is consistent with the synthetic having no real vol-regime alpha. The intervention point matters; whether the alpha exists is a separate question that needs real data to answer.