What 19 arms told us about strategy composition

The harness started as a 3-arm A/B (baseline + blackout + drift) and grew to 19 arms across two synthetic signal modes (single-signal: FOMC drift only; dual-signal: FOMC drift + short-clock mean-reversion). Across that growth, four claims emerged that survived multi-seed measurement. This note names them — and the discipline rules they support — so a future operator can find the evidence without paging through every PR.

Claim 1: Strategy shape matters more than factor count

time_series_momentum (one factor + 2% threshold) beat four_factor (four moments combined) on mean AND on min Sharpe. The threshold doesn’t add information — it SUBTRACTS trading when the signal is weak. Cross-sectional strategies always trade the quantiles; the entry-rule axis closes a chop hole the score axis can’t.

→ Documented: strategy-shape-beats-factor-count

Catalog impact: the entry-rule axis (xs_momentum_with_spread_filter, time_series_momentum, xs_momentum_with_vol_regime_filter) is now an explicit row in the decomposition table.

Claim 2: Composite-horizon beats single-window

xs_three_clock_momentum (5-bar / 20-bar / 60-bar linear combination) posted +1.429 mean Sharpe vs baseline’s +1.001 — a 0.43 improvement, second only to the vol-regime gate (+1.493). The negative coefficient on mom_5 (the short window) is what distinguishes “composite” from “averaged” — short-window mean-reversion + medium-window trend stacks on the same strategy.

→ Documented: three-clock-momentum-tops-the-harness

Catalog impact: a “Multi-window single-factor” row in the score-axis table.

Claim 3: Mean Sharpe rankings hide the variance trade

vol_regime_filter won on mean Sharpe (+1.493) but came with stdev 1.383 — more than 2× baseline’s 0.581. The min went negative (−0.069). Ranking strategies by mean alone misses the operator-relevant trade.

→ Documented: vol-regime-filter-mean-vs-variance

ADR impact: ADR-0060 (“Harness measurement protocol”) codifies three rules — report all four stats, rank by operator metric, run composite arms before declaring two top arms.

Claim 4: Composite stack-vs-interfere depends on the data

Composing the two top arms (three_clock_momentum + vol_regime_filter) produced an interesting test:

• On single-signal data: composite mean +1.460 SAT BETWEEN parents. Min −0.310 (worse than either). INTERFERES.

• On dual-signal data: composite mean +1.190 ABOVE both parents. STACKS.

The strategies didn’t change. The data changed. Both arms work via noise-reduction on single-signal data (they smooth the same noise); on dual-signal data, they catch different signals.

→ Documented: composite-strategies-can-interfere (the interfere result) + dual-signal-makes-composites-stack (the stack result).

Regression-test impact: tests/test_example_fomc_blackout_compare.py now has two tests that pin the fingerprint property of the dual-signal mode. If a future change breaks the predicted ordering, CI fails.

The five-step process

The work followed a predictable shape:

1. Add a strategy. Code change, unit tests, lookback ordering, lookback property for the build planner. PR.
2. Add to the harness. 1-tier import + arm config + print row + multi-seed dict entry. PR.
3. Read the numbers. Multi-seed stdev / min / max produce evidence. PR description names the observation.
4. Generalise. Research note articulates the discipline rule the observation supports. PR.
5. Codify if load-bearing. ADR or regression test pins the rule against future drift. PR.

The pattern doesn’t have to run in order — Claim 1 was named first (PR #557) but the entry-rule axis it predicted didn’t exist as a strategy until PR #561. Claim 4 ran the predicted experiment three PRs after the question was asked. The research note + code + test triple is what survives; the order of the triple within a question is incidental.

What this session DIDN’T show

• Real-data results. Every claim is on synthetic data designed to surface specific edges. Real markets have many more signals (and many more noise sources). ADR-0014 (stale-thesis policy) says claims need real-data verification before deployment — this work is the scaffold for that verification, not the verification itself.

• The harness’s sensitivity to seed count. All claims here ran at N=5 seeds. A future operator can re-run at N=20 or N=100 to refine the stdev / min estimates. The 4-stat reporting doesn’t change; just the precision does.

• Cross-strategy correlation. The harness reports per-arm stats, not pairwise correlation. Composite arms test pairwise interaction; a fuller correlation matrix would name more shape-vs-shape relationships. Out of scope.

The discipline-rule index

Three rules from ADR-0060 + four claims above:

#	Rule	Source
R1	Report all four (mean, stdev, min, max)	ADR-0060
R2	Equal-leverage controls for Sharpe comparisons	ADR-0060
R3	Run composite arm before declaring two top arms	ADR-0060
R4	Entry rule does more work than score rule on most signals	Claim 1
R5	Multi-window single-factor beats single-window	Claim 2
R6	Higher mean ≠ better choice; min Sharpe is the operator gate	Claim 3
R7	Composite stack-vs-interfere is a property of data, not strategies	Claim 4

The next experiments

Out of scope for this session, named for future-me:

• A third synthetic mode (vol-clustering autocorrelation) to test whether the realized_vol_of_vol_20 and vol_ratio_5_60 features unlock new edges.
• Pairwise correlation matrix across the 19 arms — which arms are catching the same signal vs. different signals?
• Real-data run: point the harness at SP500 + Polygon daily bars + the live FOMC calendar. Most of the synthetic claims will either confirm, refute, or refine — but the harness is the test surface that can run that comparison.

The 19-arm A/B + the discipline rules + the regression tests are the platform that makes the next experiments cheap to attempt.