statarb-liquid-cap-v1 — a liquid, position-capped residual-reversion book (S-053 RELEASED)

ID: none — S-053 released 2026-07-15 (owner). Dropped exploration; the number is free for the next registration (it was the highest, so no gap). Kept as an honest record of the work, un-numbered. Slug: statarb-liquid-cap-v1 Type: Stat-arb pillar — tradeable variant of the frozen S-038 (D-008) residual-reversion book. Date: 2026-07-15 Status: DROPPED / not pursued — un-numbered exploration, NOT frozen, NOT counted, NOT live, no forward paper. The robustness read (below) is why: it solves the tradeability constraints but is a lumpy, regime-driven strategy (full Sharpe 0.44, −40% max drawdown, negative in a third of the years; recent OOS carried by 2025-26), not a clean market-neutral carry. Not worth a live number. Motivation: the owner's two constraints on the working S-038 book: (1) too many positions (~188) and (2) some illiquid names. S-052 showed the edge can't be moved to a few ETFs (sector level = momentum, not reversion). This keeps the real stock-level reversion edge but makes it tradeable.

Idea — two existing knobs, no engine edit

The S-038 universe is already ranked by dollar volume (_eligibility: dvol = sum(close*volume)), so "illiquid names" are just the tail of the top-500. Two frozen-engine parameters solve both constraints: - topn — liquidity depth: keep only the top-N most-liquid names (500 → 250 → 150 → 100). Directly drops the illiquid tail the owner dislikes (and likely removes noise + cost). - max_side — cap positions per side (the strongest |s| opens kept).

Sweep topn × max_side with the otherwise-frozen S-038 config (kappa-sizing + sector tilt, clean stock universe, 10bps/side); measure Sharpe / net-yr / beta / avg #names / turnover, in-sample (≤2022) and on the 2023+ holdout. Goal: the most-liquid, fewest-position config that keeps a real, market-neutral edge. (liq_topn/max_side default to the frozen values → S-038's confirmed result is reproduced when unfiltered.)

Method

Engine: the exploration workspace (reuses the frozen statarb._backtest). Baseline context (from the max_side sweep): full book ~188 names, OOS Sharpe 1.00, beta ~0; capping to ~50/side (~100 names) held OOS 0.86; below ~40 names it degraded and got noisy (breadth matters).

Finding (2026-07-15) — restricting to liquid names HELPS (but the OOS is regime-flattered)

topn × max_side sweep (frozen S-038, IS ≤2022 / OOS 2023+):

• Dropping the illiquid tail is a benefit, not a cost: keeping only the most-liquid names (lower topn) holds/improves the OOS Sharpe (1.00 → 1.21 at topn=150), keeps beta ≈ 0, and cuts names 188 → ~56. The illiquid tail (ranks 100–500) was dragging the book down. max_side on top barely matters once topn is small.
• Honest caveat — the OOS is regime-flattered. The in-sample Sharpe is weak everywhere and falls to ≈0 at topn=100; the whole OOS strength is a 2023+ phenomenon (liquid large-caps carried reversion recently). So do NOT chase topn=100's 1.34 (IS ≈ 0 = OOS-fishing, the S-050 trap).
• Recommended config: topn=150 (≈56 liquid names, no hard cap needed). Chosen on the liquidity rationale (drop the illiquid tail), not by fishing the best OOS: OOS 1.21, beta 0.03, IS 0.29 (weak but not negative), ~⅓ the positions, all liquid. It won't reach a handful — the edge is breadth-dependent — but it solves both owner constraints (fewer positions + liquid only) without abandoning the working edge.

Where next: because IS is weak and OOS is recent-regime, the honest confirmation is forward paper of the ~56-name liquid book (topn=150), tracked market-neutral (beta ≈ 0). Owner to direct.

Robustness — the sobering deeper read (2026-07-15), before any forward paper

the engine — the topn=150 book beyond the headline Sharpe (net 10bps/side, full 2005-2026): - Full-period Sharpe just 0.44, net +6.8%/yr, but max drawdown −40% (worst day −6.4%, worst month ≈ −15%). A −40% drawdown on a book billed "market-neutral" is a serious flag. - Year-to-year it is wildly inconsistent: strong (2008 +1.3, 2014 +1.9, 2016 +1.3, 2020 +1.3, 2025 +2.5, 2026 +3.0 Sharpe) but badly negative in many years (2006, 2012, 2017 −1.6, 2018 −1.9, 2022 −1.0, 2024 −0.6). The flattering "OOS 1.21" is really carried by 2025-2026 (two exceptional years); 2024 actually lost money. - Cost-sensitive (turnover ~48/yr): Sharpe 0.60 → 0.44 → 0.29 → 0.13 at 5 / 10 / 15 / 20 bps a side. If real costs run above 10 bps the edge thins fast.

Honest reframe: the liquid-restriction does solve the owner's tradeability problem (fewer, all-liquid names, beta ≈ 0) and does not cost OOS Sharpe vs the full book. But the deeper read shows this is not a clean, durable market-neutral earner — it's lumpy, has a −40% drawdown, is negative in a third of the years, and the recent strength is a 2-year regime. The headline Sharpe hid all of that. This is a conditional/regime strategy, not a steady carry — size and expectations accordingly.

Forward-paper artifact (2026-07-15) — the book is ready

the engine emits today's target book from the frozen engine at topn=150. As of 2026-05-27: 43 long + 19 short = 62 liquid stock legs (all top-150-by-dollar-volume — HOOD/BABA/TMO/ISRG/JPM/AVGO long; LLY/TSLA/AAPL short …) + a 10-ETF beta-hedge basket (SPY short ~98% + sector tilts) to zero the net factor beta. Forward-paper protocol: each session, rebalance to this book (re-run the engine at topn=150), track net P&L + realized beta (target ≈ 0); the as-of date advances with each ingest. Production runner = the remaining build (a daily job that snapshots the book + records P&L), owner to green-light.

Honest caveats (known up front)

• Reversion needs breadth. Prior sweep shows the edge erodes fast below ~100 names; a shallow-liquid + hard-capped book may keep less of the OOS Sharpe. Expect a trade-off, not a free lunch. This won't reach a handful of positions — that isn't achievable with this edge (S-052 confirmed).
• Low-N configs that "look good OOS" are the S-050 small-sample trap; judge on IS+OOS consistency, not a lone OOS.
• Restricting topn is a form of universe selection — pick the depth on the liquidity rationale (drop the illiquid tail), not by fishing for the best backtest.

Where next (owner to direct)

Pick the sweet-spot (most-liquid + fewest-position that keeps the edge and beta≈0); then forward-paper it. No promotion / live cell until it clears a clean read.

Gate — [CALIBRATE] (not set)

Not set. A tradeable version needs OOS Sharpe retained vs the full book, |beta| ≤ [CALIBRATE], all-liquid names, and a position count the owner will actually trade.

ONGOING exploration; the frozen S-038 (D-008) engine is reused unchanged. Do NOT count or freeze until go.