Momentum 12−1 long-only v1 preregistration

Ledger: trial #21 preregister event rule hash bcf45aad640b041e17f5660225976e20dacc875c536f57f86248959c948ff796 · Screening window 1998-01-01→2021-12-31 · Lockbox 2022-01-01→2026-08-21 (opened once, by unlock_lockbox event)

LOCKBOX STATUS AT REGISTRATION (2026-08-27T00:15Z): UNOPENED. The 2022-01-01→2026-08-21 lockbox has not been inspected, computed, printed, or used in this document. Every performance or capacity number below comes only from the screening-window momentum-positive confirm-card.

1. Mechanism (who pays, and why they keep paying)

The proposed payer is a persistent combination of behavioural underreaction and institutional constraints: investors incorporate slow-moving information incompletely, while benchmark, mandate, career-risk, and trading constraints delay the capital that would immediately arbitrage cross-sectional winners. The effect should appear in the liquid top-ranked decile after skipping the most recent month, where the signal avoids short-horizon reversal; it should weaken when abrupt reversals dominate or when crowded winners gap before the lagged sizing rule can react. The negative control is the bottom-decile portfolio: it must not outperform the registered top-decile portfolio, because that would reverse the mechanism's predicted ordering.

2. Exact rule (frozen)

• Universe: US common stocks, unadjusted price ≥ $10 and 20-day ADV ≥ $10M at formation; classification-free. TICKERS size/category/industry metadata is current-only and may be used only for diagnostics and the cost bucket, never for eligibility.
• Signal: at each month-end formation close, closeadj(prior month-end) / closeadj(12 month-ends prior) − 1. Both anchors are adjusted closes known by formation close; the most recent month is skipped.
• Portfolio: long-only equal-weight top decile of eligible signals; rebalance monthly at month-end. Decide at the formation close, fill at the next session adjusted open, and exit at the next rebalance adjusted open. A delisted name exits at its last observed close.
• Costs: charge 10 bps per side for current-metadata Mid+ names and 25 bps per side otherwise, including initial, rebalance, terminal, and exposure-change trades. The buckets are not point-in-time and can misclassify historical size; this is a frozen modelling caveat, not an eligibility filter.
• Sizing overlay: use the prior 60 completed unscaled net daily returns, with a one-session lag, to target 10% annualized volatility; cap leverage at 1.0. The overlay is included, not optional, and cannot pre-empt an overnight gap.
• Parameters: k1 = 12 month-end lookback, k2 = 1 skipped month; top-decile cut, monthly rebalance, trailing-60-day volatility window, 10% volatility target, and 1.0 leverage cap. These are the confirm-card candidate settings and are frozen before the lockbox opens; no post-lockbox plateau selection is permitted.
• Excluded: no sector, industry, current-size, quality, data_sane, earnings-window, or volume-shock filter; no optional variant remains. Missing required formation inputs make a name ineligible for that formation only.

3. Predicted shape (from the screening window — the confirm-card numbers)

• Primary unscaled result: net excess versus the equal-weight universe was 4.87%/yr with Newey–West t = 1.25. The long-only portfolio returned 15.0%/yr net, Sharpe 0.49, with max drawdown −68.5%; mean one-way turnover was 37.6% per month.
• Included sizing rule and trade-off: the trailing-60-day, 10%-target, 1.0-cap overlay lowers the registered screening expectation to 6.8%/yr net while improving Sharpe to 0.55 and max drawdown to −27.6% (reported here as −28% rounded), versus 15.0%/yr, Sharpe 0.49, and −68.5% (−68% rounded) for the unscaled fundable portfolio. The reduction in return is the price of the drawdown reduction; the overlay was developed in-sample and cannot prevent gaps.
• Costs and jitter: the unscaled portfolio remained at 13.9%/yr under 2× costs and 12.7%/yr under 3× costs. Moving the decision close and both signal anchors by one trading session, while still filling at the following open, produced 13.3%/yr at −1 session and 13.2%/yr at +1 session.
• Random twins: all 100 uniformly randomized formation-cross-section twins trailed the actual portfolio; the empirical one-sided probability was p = 0.0099 for both absolute annual net return and active return versus the equal-weight universe.
• Right-tail dependence: removing the best 14 months (the best 5% of months) left only 0.41%/yr and Sharpe 0.02. The screening edge is therefore strongly dependent on a small right tail; this is a central fragility, not a robustness success.
• Capacity: median position 20-day ADV was $36.7m. The median equal-weight binding sleeve at 1% of ADV was approximately $14.9m.
• Current-metadata size diagnostic: Large/Mid returned 21.5%/yr with Sharpe 0.73, while Small/Micro returned −24.5%/yr with Sharpe −0.56. Because the classification is current rather than point-in-time, this is diagnostic only. The supplied confirm-card did not report screening sub-period estimates, so none are invented here.

The table is the confirm-card's screening-window calendar-compound net series for the unscaled fundable portfolio; 1998 is unavailable because the first exit occurs after the required lookback.

4. Falsifier (pre-committed lockbox gates — any miss = REFUSED, no "fix and retry")

• Lockbox net excess versus the equal-weight universe must be > 0 and at least 50% of the screening-window 4.87%/yr estimate.
• Newey–West t must be ≥ 2 on the lockbox alone, and at least 3 of the 5 lockbox calendar years must have positive absolute net return.
• At 2× the frozen per-side costs, lockbox net excess must remain positive.
• Removing the best 5% of lockbox periods must not reduce lockbox annualized net return below −5%/yr. This deliberately tolerates some right-tail dependence but refuses a lockbox carried by an extreme tail whose removal breaches that floor.
• The ±1-session jitter variants must each remain within 25% of the lockbox headline net excess. They are diagnostics against timing luck, not alternative rules.
• Deflated Sharpe Ratio against the full trial ledger must be ≥ 0.95. Raw N = 21 (15 free looks + 5 confirm-cards + this preregistration); clustered effective N = 21, conservatively applying no correlation discount because no clustering receipt exists. Both counts and the no-discount assumption must be reported with the one-shot verdict.
• Negative control: the bottom decile must not outperform the registered top-decile portfolio in the lockbox.
• These gates are conjunctive. Any miss yields REFUSED; the lockbox may not be repaired, re-specified, or reopened for a retry.

5. Live contract (if it passes)

• Stage and venue: paper_promotion → one-week Alpaca live-path dress rehearsal on 8–12 Sep 2026 → live_graduation on 15 Sep 2026. A passing candidate then begins tiny Alpaca live trading as incubation, not as evidence that full size is safe.
• Candidate set and sleeve: admit up to 5 survivors, long-only, inside a combined sleeve ≤ £10k.
• Sizing: start each strategy at 10 bp of NAV, subject to the combined sleeve limit, the included volatility overlay, a 1% ADV cap, and whole shares. Scaling 10→25→50 bp is allowed only through separate sealed approvals; no stage is automatic.
• Kill: live-risk-policy/v1 must be bound by hash in the graduation receipt before any live order. Hard-flat on strategy day loss > 1.25× simulated 99.5% loss, drawdown > 1.25× simulated 95th-percentile max drawdown, or any reconciliation/data failure. A path below the 5th-percentile Monte Carlo envelope is a soft breach and cuts size to 25%.
• Review cadence: weekly for the first month; retire after two consecutive soft-breach reviews. Hard breaches do not wait for review.

6. Trial accounting

• Source free-look trial #1: momentum-12-1, with one allowed variant; confirm-card id: momentum-positive; outcome: this momentum-12-1-long-only-v1 pre-registration at ledger trial #21.
• Trial-count basis: 15 free looks + 5 confirm-cards + this pre-registration = raw N = 21; clustered effective N = 21 under the conservative no-discount rule recorded in §4.
• This pre-registration's ledger rule hash is bcf45aad640b041e17f5660225976e20dacc875c536f57f86248959c948ff796. Every lockbox calculation, dress-rehearsal receipt, risk-policy binding, live decision, scaling approval, and later run must log against momentum-12-1-long-only-v1.

Amendment v1.1 — overlay incubation (sealed before lockbox opening)

Amendment date: 2026-08-27 · Owner decision: Q12=A

Before 60 live sessions exist, the 60-session overlay window is filled with the strategy's OWN simulated unscaled net daily returns for the sessions preceding the arm date, produced by the identical kernel on sealed Sharadar data via the live-lane simulate-window command, sealed as simulated_daily_return receipts hash-bound to the refresh manifest + policy sha + strategy spec hash. Live sessions replace simulated ones as they accrue. The window is computable only after a sealed lockbox_opened event. There is NO screening-window seed and NO floor fallback. Without the complete sealed window, the lane refuses to trade.

This amendment is sealed BEFORE the lockbox-opening event, planned ≤ 2026-09-03. It does not alter §2's rule text or the strategy rule hash bcf45aad640b041e17f5660225976e20dacc875c536f57f86248959c948ff796.

The companion preregistration event binds this amended document without pretending a self-referential file can hash its own printed digest. Hash the final Markdown bytes after replacing only the next line with the same label and no value; the newline remains.

• Preregistration document SHA-256 (self-reference line blanked): 5bd70065a19312c5db0dd2e55b6c38288883fd473741fbb1583d42c74e1d3e8a → 84198caeaf294eabc4edcd4596497a68a3e471ac631db5f46f2d7a04a0c6751b
• Verification: perl -0pe 's/^- Preregistration document SHA-256 \(self-reference line blanked\):.*$/- Preregistration document SHA-256 (self-reference line blanked):/m' docs/research/prereg-momentum-12-1-long-only-v1-2026-08-27.md | shasum -a 256