"Edge-discovery batch 5 — option-market proxies (dividend / low-vol quality)"

Date: 2026-09-26 | panel/split/costs as before. Candidates: H15 dividend-timing momentum (trailing cash dividend yield x trailing momentum; dividend clientele is a well-documented option-market clientele — low-carry names carry richer calls / put skew), H16 quality-tilted low-vol (low realized vol interacted with the only fresh PIT quarterly field available locally, the earnings surprise z). Q0 = plain low-vol baseline (no promotion; H16's p_vs_base is the quality increment over the low-vol core). DSR N=17 cumulative. Placebo = within-date permutation (200 draws), gate at p<0.05.

#Scoreboard

#H15 dividend-timing-momentum -> RESEARCH

• OOS Sharpe 0.81 (IS 0.44), p vs 0 0.01, p vs Q0 0.15, corrQ0 0.68, beta 0.90, placebo p 0.48 (null mean 0.81)
• v2: DSR nan (N=17), PIT-200 0.77, lag2 0.83, tiered 0.78, pos-years 11/11
• cost ladder: optimistic=0.83, base=0.81, conservative=0.74, severe=0.62
• decay: H5=0.81; H10=0.75; H21=0.81; H42=0.84; H63=0.90
• yearly: 2016:2.08; 2017:1.47; 2018:0.15; 2019:1.42; 2020:0.03; 2021:1.62; 2022:0.30; 2023:0.64; 2024:1.36; 2025:0.97; 2026:1.55
• score parts: {'oos_sharpe': 10.1, 'signif': 14.9, 'regimes': 15.0, 'costs': 7.7, 'perturb': 10.0, 'breadth': 5.0, 'capacity': 4.9, 'decorr': 1.6, 'plaus': 2.2}

#H16 quality-low-vol -> RESEARCH

• OOS Sharpe 0.59 (IS 0.72), p vs 0 0.03, p vs Q0 0.93, corrQ0 0.86, beta 0.56, placebo p 1.00 (null mean 0.73)
• v2: DSR nan (N=17), PIT-200 0.69, lag2 0.58, tiered 0.55, pos-years 10/11
• cost ladder: optimistic=0.62, base=0.59, conservative=0.51, severe=0.37
• decay: H5=0.66; H10=0.66; H21=0.59; H42=0.63; H63=0.65
• yearly: 2016:0.55; 2017:2.22; 2018:0.53; 2019:2.30; 2020:-0.81; 2021:1.68; 2022:0.03; 2023:0.87; 2024:1.75; 2025:0.01; 2026:2.15
• score parts: {'oos_sharpe': 7.4, 'signif': 14.6, 'regimes': 15.0, 'costs': 6.2, 'perturb': 10.0, 'breadth': 5.0, 'capacity': 4.9, 'decorr': 0.7, 'plaus': 2.5}

#Q0 plain low-vol baseline (no promotion)

• OOS Sharpe 0.79; exists only to measure H16's quality tilt increment.

#Reviewer note

• Book-to-market is deliberately absent: the _ADJ source is split/dividend adjusted and the only unadjusted fields cached are open/close, so a value factor cannot be built PIT from local free data. H16's profitability proxy is the earnings surprise z (the one fresh PIT quarterly field in the earnings parquet). H15's dividend signal uses the dividend column already inside _ADJ.
• H16 is long-only low-vol by construction; its correlation with Q0 is high, so the raw OOS Sharpe is NOT evidence — the p vs Q0 increment is. Both candidates are short-sample caveats.

#Reviewer addendum

1. H15 is dividend yield x momentum, and momentum is doing the work. corr vs Q0 0.68; the placebo says the cross-sectional ranking does not beat a random shuffle (p 0.48, null mean 0.81 == observed 0.81). IS 0.44 < OOS 0.81 is also the familiar OOS-tilt warning.
2. H16's quality tilt is neutral-to-harmful. Long-only low-vol core Q0 = 0.79; adding the surprise-z quality tilt lowers it to 0.59 with p vs Q0 0.93 (a wash at best) and placebo p 1.00. The surprise z is a weak stand-in for a real profitability measure — a data-limited null, not a clean kill of the quality-low-vol mechanism. Reopen only with a genuine fundamentals feed (gross profit / assets), not a re-grid of this proxy.
3. What this batch actually bought: the dividend column plumbing in load_panel and the PIT-quarterly quality_panel in event_signals are reusable for a future fundamentals-backed value/quality study. No candidate reaches CANDIDATE; nothing is wired to any sleeve.

Runtime: 14.4 min.