MK3 regime-design inputs: lead-lag study + re-entry churn

Two design questions surfaced 2026-05-15 that belong in the MK3 spike, not MK2. Both are regime-detection inputs. Neither is a Codex script - they are analyses to run during MK3 design against the corpus + the purchased 1-year UW bundle.

1. Lead-lag / predictive-power study (calibrates the “early and right” mandate)

User hypothesis under test: “Tide leads price.” Held as a directional prior, explicitly to be tested, not assumed. The engine already consumes Market Tide as the Premium Flow pillar (14 pf_* features per scoring_event: pf_tide_dir, pf_short_tide_dir, pf_tide_roc, pf_net_prem_dir, pf_net_prem_accel_dir, pf_intraday_reversal_dir, pf_reversal_mode, …). The data is captured; what is NOT established is whether Tide actually leads price, and whether it leads better than the alternatives.

The study. Measure lead-lag / forward-predictive power of each candidate leading signal against forward SPY and forward suggested-option returns at multiple horizons (e.g. 1/3/5/15/30 min), across regimes:

CandidateSource
Market Tide (pf_tide_dir, pf_net_prem_accel_dir, short-vs-day divergence)scoring_events
IMPULSE engine conviction (by trigger: hiro / strike_stack / compound)scoring_events.trigger + feature_json
Composite scorescoring_events.score
Dealer / gamma positioning (dp_*)scoring_events
Raw price momentum / VWAP-extensionsession_features

Method: cross-correlation lead-lag + conditional forward-return / predictive feature importance. Run on (a) the live corpus (with #103 counterfactual backfill + 109 canonical P&L now clean) and (b) the 1-year historical UW bundle (the regime-spread backtest already scoped in the mk3-build-week plan).

Output: which signal leads, by how many seconds/minutes, and whether the lead is stable across regimes. This is the empirical calibration of the “early and right” mandate - “early” must be anchored to whatever the study shows actually leads. Today’s corpus work already shows IMPULSE ≈ baseline except hiro+strike_stack, so “Impulse leads” is also unconfirmed. Settle it with measurement.

Worked example to include (today’s losses): BULL entries 09:35 and 09:58 fired at score 67-68 with pf_net_prem_accel_dir = -1.0 (Tide net-premium accelerating DOWN - the UW guide’s explicit bearish criterion) → both lost. Aggregate today: 24 BULL entries (avg pf_tide_dir +0.27, weak) vs 7 BEAR (avg -0.39, strong) while the day’s tide was net bearish (NCP -271M). The composite has the Tide signal and underweights it (structural Dealer/Vol BULL bias drowns it). If the study confirms Tide leads, the design implication is a Tide-divergence regime veto: when composite bias opposes Tide-acceleration, skip or fade.

2. Re-entry / churn vs single-hold (over-trading)

Question: on a clustered signal, does rapid re-entry destroy value vs. taking one position and holding it through the cluster?

Worked example: 2026-05-15 trades #509 (740C, +0 scratch, 12:06), #510 (741C, +400, 12:09), #511 (741C, open/losing, 12:12) - three re-entries on the same BULL thesis in 6 minutes, each trail-exiting near scratch while the underlying move (if any) was chopped into fragments. Same over-trading concern the user has raised repeatedly (“fewer trades, just the winners”).

The analysis (MK3 design, data already present): reconstruct “single position held from first cluster entry through last cluster exit” vs “the actual N-trade churn” from trade_path_snapshots + signal-cluster timing. Slots next to the existing findings on 2026-05-14-five-corpus-rules-for-mk3: post-loss cooldown (after 1 loss WR 65%→49%), hold-time (<2min 76% WR vs 10-20min 37%). Re-entry churn is the same family: a debounce / position-hold-policy decision for MK3.

Why both are MK3, not MK2

Signal/strategy changes bias the live corpus and burn Codex cycles on a system being replaced (the week’s standing discipline: only data-integrity gets MK2 patches; signal/regime → MK3). The data to answer both already exists in the corpus and the 1Y bundle. These are spike design inputs, not scripts.