Backtest: Retest — re-entry ladder evidence

1. The full cycle table — before / during / after each retest

Phases: Rebound = running low → +10% rebound high · Pullback = rebound high → retest trigger · After = 63 sessions post-trigger.

[rate-shock veto] = 10Y rose during pullback · [OFFLEAD] = offensives leading at trigger (s20>0 or HB/LV > 50d SMA)

L5b round-trips — 12 stopped cycles

Realized losses range from −1.2% to −15.2%, mean −8.3%. The stop converts the census's average −17.2% failed-cycle excursion into bounded realized losses and leaves L5b 0% deployed at both secular ultimate lows. Gap-throughs remain possible: the 2008-11-05 cycle realized −15.2% despite a nominal −5% stop level.

Named episodes

2. The anatomy of a healthy vs failed retest cycle

The normal retest is a growth scare: yields fall during the pullback, defensives lead, then offensives resume after a successful hold. The failure signature is when the pullback is rate-driven — yields rising into the retest means the drawdown's cause (tightening/inflation) is still active. Rebound-phase and pullback-phase HB/LV direction are nearly identical in both groups — only the persistence of offensive leadership into the trigger day ( [OFFLEAD]) carries signal.

3. Gate 1 — rate-shock veto (skip if 10Y rose during pullback)

• Vetoes 2008-06 (+24bp → −32.6% fwd 6m), 2022-04 (+33bp → −18.0%), 2018-10, 2022-08 — all genuine rate-shock regimes. Strongest single linear relation in the study: pullback Δ10Y vs fwd 6m, r = −0.45.
• Honest cost: vetoes 2 winners — 2000-01 (+22bp, +1.4% fwd 12m, mediocre) and 2024-04 (+42bp, +18.9% fwd 6m, zero drawdown — a clean winner). The gate fails when a rate shock is small, self-limiting, and growth stays intact. No in-sample separator found (rate level / CPI regime is a hypothesis, not a tested rule).
• The gate improves forward returns, not per-trade path risk (MAE −4.9% vs −5.2% — no separation).
• Threshold robustness: >0 / >+10 / >+20bp all veto 5–6 trades with identical direction; fwd 6m kept ≈ +11% vs vetoed −3.4% to −5.6%.

4. Gate 2 — OFFLEAD sizing tilt (offensives still leading at trigger)

• The 3 stacked signals are the three generational retests: 2002-11-11, 2009-03-30, 2011-11-01 — fwd 12m +21.6 / +52.2 / +18.7%, two with zero post-entry drawdown. Offensive leadership persisting through a retest marks the true cycle turn.
• Conflict cases (both flags): 2000-01 (held, mediocre) and 2022-08 (stopped) — when the two disagree, the rate gate was right both times; OFFLEAD must never override the veto. Tilt, don't veto: 100% size when on, reduced when off.

5. Relationship to trough depth

• Stopped triggers sit in far deeper episodes (mean ep DD −36.0% vs −24.4% for held). All 8 triggers that fired before their episode's ultimate trough were stopped mechanically — the anchor is, by construction, above the final low.
• Of those 8 doomed pre-trough trades, the rate gate flags only 2 (2008-06, 2022-08). The other 6 were growth-scare failures with falling yields (dot-com −11/−19/−17bp, GFC Oct–Dec −3/−2/−19bp) — invisible to both gates ex-ante. The stop-loss remains the primary defense; gates are a return-enhancer, not a crash-shield.
• Deep episodes (>25% DD): 10 triggers, 3 held, only 1 rate-vetoed — deep bears are growth-driven, yields fall in pullbacks. Shallow episodes (<15%): 7 triggers, 3 held, 4 rate-vetoed — shallow drawdowns in hiking/inflation scares are exactly where the rate gate concentrates.
• HB/LV pullback behavior has no relation to trough depth (r ≈ +0.07); rate-rise-during-pullback correlates mildly with shallower ultimate depth (r = +0.13) — rate shocks produce sharp but shorter drawdowns.

7. Magnitude dose-response & sizing simulation (2026-11-03, owner question)

7.1 Does magnitude matter? Yes — at the extremes, and non-linearly

10Y pullback move buckets:

Dates in the 6/6 bucket: 1998-10, 2010-08, 2011-11, 2016-06, 2019-05, 2020-04. The middle bucket is polluted by slow-bleed secular mid-retests (2001, 2002×2, 2008×3, 2016-01) — a modest yield fall is the signature of grinding bears, not capitulation. Within rate-kept trades, corr(Δ10Y, fwd 6m) = −0.42 (bigger fall → better). Do not size linearly in Δ10Y.

Offensive-lead magnitude: strong (s20 > +10pp) = 3/3 held (2002-11, 2009-03, 2011-11; fwd 6m +20.1%, MAE −1.6%); marginal flag (HB/LV>SMA50 only, s20<0) = 1/2 — and both marginals are the rate-gate conflicts anyway.

The union — "strong" = 10Y fell >20bp OR s20 > +10pp: 8/8 held (adds 2002-11, 2009-03 to the six above), fwd 6m +23.1%, fwd 12m +29.0%, worst MAE −4.7%. Complement: 3/12 held, fwd 6m −1.0%. Binomial p ≈ 0.0027 — the first split in this research program to touch the Bonferroni bar (~0.005). Caveat: the union is itself a post-hoc composite, so effective multiplicity is worse than the raw bar.

7.2 Owner's proposed sizing rule, simulated through the canonical engine

Rule as proposed: at each L5b trigger — 0% if 10Y rose during pullback; 35% if 10Y fell; +35% (→70%) if offensives also lead (s20>0 or HB/LV>50d SMA); residual stays in 2Y cash. Also tested a magnitude-tiered form (70% tier = "strong" union above) and a 0/50/100 scaling.

*fwd per deployed dollar is identical across sized variants by construction (relative weights coincide within multi-buy episodes; single-buy weights cancel). **Chained here = episodes compounded with 2Y cash between; pure L5b reads 1.80x vs the report's 1.89x (minor chain-method detail, identical across variants — comparisons valid).

7.3 Where the user's scheme wins and pays (per-episode ep_ret vs pure L5b)

• Gains, all in secular/extended bears: dot-com +21.4pp (−20.0% → +1.3%), GFC +18.2pp (−0.7% → +17.4%), 2022 +7.3pp. Mechanism: stopped-trade capital damage falls from 0.997 to 0.247 units of episode capital across the 12 stops.
• Costs, all in V-bottom corrections: 2020 −14.5pp (+22.4% → +7.9% — only 35% deployed at the best single entry of the whole study), 1998 −5.1pp, 2010 −1.5pp. The 2020 cost is precisely what the magnitude tier fixes: 10Y fell −26bp there → "strong" → 70% under V-mag.
• The scheme is a secular-bear insurance overlay, paid for with correction upside — the same regime trade-off as every construct in this program. It is not a free lunch.

7.4 Strategist assessment

The 0/35/70 structure is sound and pre-registrable: veto first (rate), tilt second (leadership), never inverted (both conflict cases resolved in the rate gate's favor). Recommended frozen form — L6c: 0% if Δ10Y(pullback) > 0; 35% base if ≤ 0; 70% if "strong" (Δ10Y ≤ −20bp OR s20 > +10pp) — which dominates the literal rule in-sample on every axis and is defensible a priori from the dose-response evidence. Known retained costs: 2024-04-19 gets 0% (clean winner skipped); the 35% floor still bleeds in the fell-0–20bp bucket (8 of 12 historical stops). Statistical status unchanged: gates selected on the same 23 events, secular gain driven by 2 episodes, union split post-hoc — freeze now, validate out-of-sample, no further tier-shopping on this dataset. Analytical only; any allocation role requires explicit owner approval.

8. Drawdown depth at trigger — amplifier, not signal (2026-11-03)

8.1 Depth alone does NOT discriminate hit/miss

Deep triggers (<−15% below high): 5/11 held. Shallow: 6/12 held. No discrimination.

8.2 But depth AMPLIFIES signal quality — the 2×2

Reading: hit/miss is determined by the signal (8/8 vs 3/15); depth scales the payoff (+29.2% vs +17.0% within confirmed trades — deeper entry = bigger discount) and the penalty (deep without signal = serial stop-outs at depth; worst MAE −15.2%). Depth is a leverage knob on signal quality, not a gate.

8.3 Naive 3-factor count (rate-kept + union + deep) — robustness check

Score 3: 4/4 held, +29.2% · score 2: 5/11, +6.3% · score 1: 0/2 · score 0: 2/6, −3.4%. Monotone gradient, but no better than the union alone — "deep + merely-modest yield fall" (score-2 secular retests) is exactly the slow-bleed trap. Keep L6c frozen as defined; do not add a depth tier on the same 23 events.

8.4 "Secular-bear insurance, paid with correction upside" — the formal statement

L5b's anchor stop already converts every failed retest into a bounded ~−8% realized loss. Therefore:

• In corrections (17/19 episodes) the retest usually holds and the market V-recovers; the only risk is under-deployment. Any sizing gate's cost here = foregone upside on the undeployed fraction — bounded, partial, and incurred in a rising market (2020: −14.5pp ep-ret from 35% vs 100% at the best entry; 1998 −5.1pp; 2010 −1.5pp).
• In secular/extended bears the retest fails repeatedly (8 of 12 stops cluster in dot-com/GFC/2022); serial full-size entries bleed ~−8% each. The gates' vetoes/underweights concentrate exactly there (+21.4pp dot-com, +18.2pp GFC, +7.3pp 2022; stopped-trade capital damage 1.00 → 0.25 units; chained 1.80x → 2.40x).
• Asymmetry: the premium (correction upside foregone) is bounded per episode (you still hold 35–70% of a recovering market), while the payout (secular serial-bleed avoided) is unbounded in repetition. Since regime is unknowable in real time (RSM 0/4320, rotation 0/21), the sized policy deliberately overpays in the frequent benign case to cap the rare ruinous one — structurally identical to buying drawdown insurance.