Phase 2f: Strategy Search Campaign

Phase 2f ran the owner's brief — improve H1 with meaningful structural changes (no data farming), and search for a genuinely new strategy that beats H1's Sharpe, diversifies it (negative/zero correlation), or has verified expansion potential — as two agent-orchestrated search workflows plus one focused follow-up round. Eighteen pre-registered hypotheses went through the full ledger protocol; every promising result was adversarially audited (patch re-applied in a clean worktree, numbers reproduced, diff reviewed for look-ahead, the append-only run log checked for undisclosed variants).

How the search was run

• Pre-registration before data. Every candidate's rule, grid (≤ 9 combos, most ≤ 6), and fixed priors were written into its module docstring before the first real-data run. Grids were frozen after first OOS contact; one disclosed bug-fix mulligan allowed.
• One scoring path. Every run goes through research.run_protocol — the ledger protocol (walk-forward grid re-fit per 504/126-bar window, stitched-OOS-only scoring, 1 tick/side slippage debited, moving-block bootstrap 90% Sharpe CI, same-span buy-and-hold, slippage and fill sensitivity) plus a Phase 2f addition: correlation to champion H1's own stitched OOS returns, the number the diversification criterion actually needs. Every invocation appends to an append-only runs.jsonl audited after the fact.
• Adversarial verification. Candidates claiming results were re-run from their exported patch by an independent auditor agent; every audited reproduction matched to four decimal places.
• Deep-history stress. Daily candidates that cleared their micro CIs were re-run on the ES/NQ e-mini parents (2012–2026 OOS, ~3× the micro span) — now possible offline via the committed data/fallback/ parquets (this environment's R2 token is read-only; see the storage-layer fallback added this phase).

Champion reference throughout: H1 OOS Sharpe 0.84 (MES) / 0.86 (MNQ), deep parents 0.74 (ES) / 0.58 (NQ), all CIs excluding zero.

Improve-H1 workflow — 8 structural candidates

Candidate	Mechanism	MES / MNQ (ES / NQ) OOS Sharpe	Verdict
i_stress_exit	VIX-latched exit: stress entries get prior-10-high target + 15-bar leash	1.06 / 0.76 (0.77 / 0.60)	⭐ finalist — note; registered, forward-validate on MES; not adopted (MNQ −0.10, best-of-10 selection risk)
i_dip_provenance	Size 2× when the dip formed intraday vs overnight (hourly-derived, info H1 can't see)	1.00 / 0.83	❌ shelved — MES active-stream alpha real (0.95 vs 0.84) but MNQ gain is pure leverage (active-stream 0.56 vs 0.86, maxDD −4.9%→−6.6%)
i_volume_capitulation_size	2× size on volume-spike capitulation days	0.93 / 0.88 (0.71 / 0.51)	❌ rejected — entire gain sits in the single final OOS window; ex-final-window it underperforms H1 on both micros; deep-history increment negative
i_vol_target_exit	ATR-scaled profit target anchored at entry	0.79 / 0.83	❌ CONFIRMED-negative — static vol targets are procyclically demanding (time-stops rose 26%→37%); H1's decaying prior-high exit is already implicitly vol-adaptive
i_bear_short_fade	Add symmetric short side below the 200-SMA	0.67 / 0.71	❌ shorts earned ~0 net at 1.2–1.4× the variance; the SMA lags bear→bull turns and fades genuine recoveries; shared grids let short noise poison long selection
i_atr_dip_entry	Replace N-day-low breach with ATR-normalized depth trigger	0.31 / 0.50	❌ selectivity starves the edge (35–48 trades); H1's edge is breadth of shallow dips
i_bear_dip_long	Buy dips below the 200-SMA (faster vol-scaled exit)	0.29 / 0.12 (0.24 / 0.40)	❌ bear-regime dips have positive means but 1.6–1.8× the vol — risk-adjusted thinner everywhere
i_ts_inversion_short	Short rallies only while VIX9D/VIX3M inverted	−0.19 / 0.10 (−0.22 / −0.11)	❌ front-end vol clocks mark spike-stress (V-recovery days), not grind-bears; systematically sells the wrong day

Improve verdict (round-2 judge, quoted): adopt no change; H1 stands. The below-200-SMA cell is closed on both sides; sizing overlays are attribution-trapped under walk-forward re-fit (quantity changes train-window Sharpe → parameter selection, so "sizing-only" deltas conflate mechanism with selection drift); exit engineering was the only axis that ever added value. Every same-book candidate landed MES-up/MNQ-flat-or-down — cross-symbol disagreement on a shared mechanism is the signature of noise around a real edge. Continuing would have been the small-improvement farming the brief prohibited.

New-strategy workflow — 8 mechanisms + 2 follow-ups

Candidate	Mechanism	OOS Sharpe (corr to H1)	Verdict
hourly_momentum (sleeve)	Long-only hourly Donchian — the untuned H5/H6 baseline promoted to a candidate	MES 0.47 (0.36) / MNQ 0.66, CI [0.14, 1.16] (0.28)	⭐ finalist — MNQ 50/50 blend with H1: Sharpe 0.95 vs 0.86 H1-alone; registered as hourly_momentum; MES leg unverified (CI spans 0)
n_vix_spike_reclaim	Buy first VIX down-tick after a ≥20% spike, ungated by trend	MES 0.51 CI [0.16, 1.05] (0.30) / MNQ 0.25 / ES 0.28 / NQ 0.22	❌ audit-CONFIRMED but mechanism inverted: all profit comes from above-SMA spikes (H1's own regime); below-SMA legs lose; blends are a wash
n_xasset_tsmom	Donchian trend on MCL/QO/BZ (new markets)	MCL 0.45 (−0.28) / QO 0.22 (0.02) / BZ 0.02 (−0.08)	❌ no CI clears zero; but the protocol's cold-start test windows structurally cannot express 100–200-day channels — slow TSMOM was not fairly falsified (see open questions)
n_xasset_pullback	H1's exact machinery on MCL/QO/BZ	MCL 0.66 / QO 0.19 / BZ 0.06	❌ closes the twice-flagged generality question: H1's reversion does not transport off equity indices (QO/BZ flat over 12–14y; MCL's 0.66 is one hot half-year, and 4×-history Brent says the bet is empty)
n_hourly_trend_breakout	Slower hourly channels (48–168h) + daily trend gate	0.09 / −0.02	❌ bounds the momentum lead: the intraweek edge lives at ≤ ~1 CME day; slowing channels to cut fees kills the signal
n_hourly_fast_symm	Fast symmetric hourly Donchian (long/short) + time stop	−0.32 / −0.23	❌ lost money even at zero slippage; hourly momentum is long-only-and-unstopped or nothing
n_bear_rally_fade	Short/flat rally fade below the 200-SMA	−0.34 / 0.01 (0.00)	❌ perfect complementarity (zero same-day overlap with H1 across 5.25y) — and no edge to put in the slot
n_fomc_predrift	Long K days into FOMC decisions (daily bars)	0.17 / 0.04 (ES 0.27 / NQ 0.27)	❌ direction right (PF 1.36–1.47 over 16y of events), never significant — daily bars bundle the drift with the post-2pm reaction
n_cot_flow	CFTC TFF positioning flow as directional signal	0.10 / 0.08 (ES −0.02 / NQ −0.06)	❌ COT on equity indexes now fully falsified (level as gate, flow as signal, micros + 14y parents)
overnight_drift (final round)	Trend-gated overnight hold, 16:00 ET close → 10:00 ET close	MES 0.45 CI spans 0 (0.43) / MNQ 0.83, CI [0.33, 1.36] (0.46)	⭐ finalist (MNQ) — the campaign's only CI-verified non-H1 edge: slippage-robust (0.87/0.83/0.78 at 0/1/2 ticks) after 2.5% fee drag; "always" won 25/26 train windows (the premium is unconditional). MES is fee-destroyed (gross 0.67 → net 0.45 → 0.22 at 2 ticks) — tick-to-notional geometry, not the anomaly, decides survival. 50/50 MNQ blend 0.99 vs 0.86, but the delta is not itself significant (P≈0.75) — paper-trade, never deploy the MES leg
fomc_hourly (final round)	Long K days into FOMC, exit 13:00 ET strictly before the statement	MES 0.58 CI [−0.05, 1.18] (0.04) / MNQ 0.54 CI [−0.12, 1.12] (0.13)	⭐ finalist (power-bound) — the most orthogonal stream found (corr-to-H1 ≈ 0), PF ~1.9, tiny drawdowns at 7% exposure, blends +0.10–0.14 on both symbols — but 42 events cannot exclude zero at Sharpe ~0.55. Isolating the pre-statement window roughly halved per-event noise vs the failed daily variant, vindicating its bounded conclusion. Accumulate events live/paper; ES/NQ parent hourly history would settle it

What the campaign established

1. H1 is a local optimum on its information set. Eight structural edits — new datapoints (hourly provenance, volume, VIX family), new sides, new entries, new exits — and none survives an adoption bar on both symbols. The two flagged finalists both need forward evidence, not more backtests.
2. Vol information belongs in the exit. Three entry vetoes failed (2c); the stress-exit improved 3 of 4 spans without ever worsening risk. This is the campaign's one directional mechanism finding.
3. The intraweek momentum sleeve is real on MNQ and is the portfolio's one verified diversifier. corr to H1 ≈ 0.3, own CI excludes zero, 50/50 blend +0.10 Sharpe. It was never tuned toward (it was the baseline), so selection bias is minimal — but the MES leg is unverified and fee drag (456–468 trades, ~1.3% of equity over the span) is the honest cost.
4. The diversification slot exists but is hard to fill. A disjoint-regime sleeve is reachable (bear_rally_fade: zero overlap days) and finding low-correlation activity is easy; finding low-correlation edge is the entire problem. Every 50/50 blend except hourly-momentum-on-MNQ underperformed H1 alone.
5. Closed doors (do not re-test without new data or a protocol change): below-200-SMA in either direction; COT on equity indexes; H1-style reversion on commodities; hourly reversion; hourly momentum outside the fast long-only band; daily-resolution event drift; static entry-anchored vol targets; sizing overlays under re-fit walk-forward.

Open questions the protocol itself raised

• Slow trend is untestable under cold-start windows. The 126-bar OOS windows give a 200-day channel zero live bars; commodity TSMOM at the literature's 6–12-month horizon was handicapped, not falsified. Warm-started (position-carrying) test windows are an owner-level protocol decision that would unlock the strongest remaining diversifier class (the QO sleeve made +5.3% through 2022 while H1 idled, directionally exactly the crisis-alpha shape).
• Portfolio capacity, not exit design, may be H1's real bottleneck: the stress-exit's MNQ loss came entirely from blocked re-entries during longer holds. A multi-cycle capacity study is a new hypothesis.

Infrastructure this phase leaves behind

• research.run_protocol + corr-to-champion scoring + append-only run log.
• Storage-layer local fallback for read-only R2 tokens; committed deep-history parquets: ES/NQ (2010–), QO (2010–), BZ (2012–), MCL (2021–) under data/fallback/.
• MCL/QO/BZ Market entries with sourced constants (BZ margin flagged as estimate) and per-market Databento roll rules (roll_rule="c" for thin books whose volume ranking flip-flops).
• New registered candidates: pullback_stress_exit, hourly_momentum, overnight_drift, fomc_hourly; plus the cross-asset machinery modules (xasset_pullback.py, xasset_tsmom.py) for future commodity hypotheses.

The forward-testing stack (post-hoc, selection-biased — a hypothesis, not a claim)

The final judge's independent portfolio math from the OOS return CSVs (1,636 inner-joined days): the two-symbol H1 book scores 0.92; adding the FOMC sleeve on both symbols → ~1.03; adding the MNQ overnight sleeve → ~1.10; adding the MNQ hourly-momentum sleeve → ~1.14. No individual blend delta is itself statistically significant (block bootstrap: overnight-MNQ P(Δ>0)=0.75, FOMC-MES P=0.82), and the sleeves were chosen after seeing OOS — so this stack is the forward-testing agenda, with every increment owed its own live burn-in.

Recommended next steps (the final judge's ranked plan)

1. Keep H1 as the sole live core on MES+MNQ (two-symbol book Sharpe 0.92); nothing in 18 candidates beat it or diversified it with significance. Build the live path (Phase 2e's recommendation), and run pullback_stress_exit beside it on MES as the cheap forward experiment.
2. Paper-trade the MNQ-only overnight sleeve (overnight_drift) — the campaign's strongest non-H1 result. Blend point estimate 0.86 → 0.99, delta insignificant (P=0.75): let live paper data decide. Never deploy the MES leg (fee-destroyed).
3. Run the FOMC hourly sleeve live/paper on both symbols to accumulate events (~0.02%/yr fee cost, 7% exposure, orthogonal). If buying data, ES/NQ parent hourly history is the single highest-value purchase — ~14y of events would settle FOMC significance.
4. Extend the pre-announcement-drift template to other scheduled macro events (CPI, NFP): same module pattern, public calendars, pre-registrable — the one untested hypothesis with a mechanism prior.
5. Optional third sleeve: the untuned hourly-momentum Donchian on MNQ (+0.08–0.10 blend at corr 0.28), behind the two event/session sleeves.
6. Owner call: warm-started test windows, then one slow-trend commodity candidate (QO/BZ/MCL, 100–200-day horizon) — the remaining untested mechanism class with a real diversification thesis.
7. Do not fund further same-book improvement searches or any of the closed-door mechanisms above.