signal-platform/docs/research/prod-book-universe-horizon.md

Production book × universe × horizon matrix

Status: PRE-REGISTERED — prepare / MacBook run; no production changes.
Branch: research/earnings-gap-and-sue
Runner: scripts/run_prod_book_universe_matrix.py

---

Question

How does the live production book (unchanged knobs) behave when we only vary:

1. History length used for entries (≈4y vs since 2016-07)
2. Tradable universe (prod ~505 vs 505 + PIT liquid Nasdaq/breadth)

No strategy modifications: same residual gate, 80/20 high-vol rank, GTL entry machinery, 3× ATR trail, 30d max hold, gate-reset re-entry, fill_mode=close, cost 10 bps/side, max 10, 1% risk.

---

Pre-registered arms (locked)

id	label	Entry start	Tradable universe
A	prod_4y_505	2022-07-01	Prod ~505 only
B	prod_4y_505_liquid	2022-07-01	Prod ∪ liquid top-1500
C	prod_2016_505	2016-07-01	Prod ~505 only
D	prod_2016_505_liquid	2016-07-01	Prod ∪ liquid top-1500

• End: last available bar in snapshot (no artificial end).
• 4y start chosen to align with recent Phase‑A / book baselines (~mid‑2022 → mid‑2026).
• 2016-07-01 = first full month after typical Alpaca floor (~2016-01); residual 12‑1 needs ~1y bars so first residual ranks appear mid‑2017 where feed allows.

Universe definitions

set	definition
Prod ~505	Symbols not in research_rank_only on the research snapshot (the original prod-universe copy).
Liquid top-1500	Point-in-time: among names with as-of close ≥ $5 and valid 63d median $vol, keep top 1500 by that $vol. Same definition as breadth IC research.
Prod ∪ liquid	A name may enter the book on date t if it is prod or in the liquid top-1500 at t.

Cross-sectional residual / vol / 80/20 ranks are recomputed inside each arm’s eligible candidate set that period (so breadth arms are not ranked against non-eligible thin names).

Explicit non-goals

• No sector residual, SUE, FIP filter, gap-cap, take-profit, vol-target, corr-cap
• No retune of trail / cutoff / min_rr
• Survivorship: report levels with the standard caveat; compare arms relatively

Reporting (required table)

Per arm: Sharpe, Sharpe SE (Mertens), CAGR %, max DD %, total return %, trades, win rate if available, start/end, n qualified longs. One markdown table + JSON.

No promotion rule — descriptive matrix only. Human decides whether breadth or depth changes the risk story.

---

Snapshot requirements

• Prefer MacBook deep research.sqlite after sector-resid deepen (prod names from ~2016, breadth deep, completion manifest complete=true).
• Race-guard before run.
• Sector map / sector ETFs optional (not used for ranking).

---

Results

Generated: 2026-07-19T14:07:37 · artifact reports/prod-book-universe-horizon-20260719-140737.json
Snapshot: MacBook deep research.sqlite (506 prod + breadth prices; 2.39M raw GTL candidates). Strategy knobs = live production (residual 80, 80/20 high-vol rank, ATR trail 3×, hold 30, gate-reset, fill_mode=close).

Survivorship: today's constituents backfilled. Compare arms relatively. Absolute deep CAGR/Sharpe are not deployable forecasts.

arm	universe	entries from	Sharpe	SE	CAGR %	max DD %	total ret %	trades	win %	vs SPY
A	505 only	2022-07-01	1.32	0.49	31.8	18.9	+205	374	35.6	+96.5
B	505 + liquid 1500	2022-07-01	0.14	0.50	−1.8	55.3	−7	706	29.3	+95.0
C	505 only	2016-07-01	0.88	0.31	16.7	24.4	+369	763	36.7	+257
D	505 + liquid 1500	2016-07-01	−0.06	0.32	−7.0	73.9	−52	1567	28.0	+254

Qualified longs: A 1 448 · B 6 587 · C 2 450 · D 11 551.

Read (relative only)

1. Same strategy, broader liquid universe kills the book (A→B and C→D). Sharpe collapses; DD roughly triples; win rate drops ~6–8pp; trade count ~doubles. This matches earlier breadth IC work: the production residual + high-vol package is a large-cap / prod-universe edge, not a “more names = better” edge.

2. Longer history on 505 stays positive but softer (A→C). Sharpe 1.32 → 0.88, CAGR 32% → 17%, DD 19% → 24%. Still well above the liquid-breadth arms. Levels are optimistic (survivorship); the useful message is “edge does not vanish when 2018/2020 are included,” not “expect 17% CAGR forever.”

3. Arm A vs older Phase‑A / short-window controls (~Sharpe 1.7–2.1): this matrix re-ranked on deep research.sqlite with a fixed entry start; numbers need not match prior reports row-for-row. Use this table for A–D comparisons, not for rewriting the production baseline number.

4. No production change implied. Keep the live ~505 universe. Do not broaden the tradable set to liquid Nasdaq under current knobs without a new pre-registered design (and almost certainly a different rank/tilt package).

Verdict

Descriptive matrix complete.

question	answer from this matrix
Prod book @ ~4y / 505	Positive (arm A)
Same + liquid Nasdaq	No — large degradation (arm B)
Prod book since 2016 / 505	Still positive, milder (arm C)
Same + liquid Nasdaq deep	No — worst arm (arm D)

PENDING_HUMAN only for whether to log “universe broaden under current knobs” as rejected in the main research index. Strategy knobs unchanged either way.