S-040 — statarb-balanced-lowturn-v1 (factor-balanced construction + lower turnover)

ID: S040 Slug: statarb-balanced-lowturn-v1 Type: Equity (market-neutral stat-arb) Date added: 2026-07-08 Status: exploration closed, NOT frozen (owner call 2026-07-13: "middenmoter" — a marginal operational tweak on S-038, not a new edge; energy moved elsewhere). The one usable finding — s_close = 0.25 gives a free −21% turnover — is folded into S-038 as a config note, not spun out as its own forward-paper setup. Component A (the hedge fix, the actual motivation) yielded nothing usable. Base: S-038

Sibling of S-038 (the frozen market-neutral reversion book). This file is the pre-registered design intent only; the [CALIBRATE] fields are the owner's judgement calls, to be frozen before any run. Per §2A, exploration precedes freeze — iterate freely first; freeze on explicit owner "go". Confirmation is forward paper, not a holdout re-run (the 2023+ vault is spent: D-005/006/008/009/010).

Why (motivation — from forward tracking, 2026-07)

Two structural burdens of S-038's sleeve A, both surfaced in forward tracking and confirmed by investigation:

1. Large, variable ETF hedge. The residual-reversion signal picks factor-tilted baskets, so the book carries ~180 bp/day of systematic (factor) return that must be hedged. The ETF hedge runs ~4–6.75× a leg (gross ~135% of net-liq at 20% per-side) and swings day to day (1.15×→4.75×). It is real, not a bug — shrinking it post-hoc leaks factor return (ridge λ=0.15 → 31 bp/day leak vs alpha of a few bp/day) and breaks beta≈0; the 60d factor matrix is full-rank so the hedge is unique. The only structural fix is to stop generating the exposure, i.e. build the legs factor-balanced.
2. High turnover. ~18%/day, daily market-on-close across ~180 names — automation-gated, costly, operationally heavy.

Component A — factor-balanced construction → small, stable hedge

Build the long/short legs so their net factor beta ≈ 0 by construction, so the residual ETF hedge is small and stable (target ~1× a leg, not ~5×). Candidate mechanisms — [CALIBRATE] pick one/blend, then freeze: - beta-matched leg selection (match long-leg vs short-leg factor loadings when choosing names); - factor-beta-neutral leg weighting (solve leg weights s.t. Σβ_long ≈ Σβ_short per factor); - sector/beta-bucketed construction (trade within factor buckets so exposure cancels intra-bucket).

Success target: raw net factor exposure ≪ S-038 (hedge → ~1× a leg) with net@3 Sharpe retained. [CALIBRATE] the acceptable hedge-shrink ↔ alpha-loss trade-off.

Component B — lower turnover over time → less trading

Cut the ~18%/day churn. Candidate levers — [CALIBRATE] pick, then freeze: - wider entry/exit bands (open |s| > [CALIBRATE], close |s| < [CALIBRATE]) — a no-trade band; - minimum holding period / trade cooldown ([CALIBRATE] days); - lower rebalance frequency (rebalance every [CALIBRATE] days instead of daily MOC); - per-day turnover cap ([CALIBRATE]).

Success target: materially lower turnover with acceptable Sharpe loss. NB the dispersion study (result record) showed breadth is load-bearing — so prefer holding longer / wider bands (fewer trades, same names) over dropping names (fewer positions), which the Q1 threshold test confirmed hurts. [CALIBRATE] the turnover ↔ Sharpe trade-off.

Base / lineage / scope

• Base = S-038 (Avellaneda-Lee residual reversion, clean common-stock universe, κ-sizing + sector tilt, MOC, net@3bps). Changing both the leg construction and the trading cadence makes this a new object → new slug/ID (not an edit of S-038).
• Scope for v1 = the neutral sleeve A only. Momentum sleeve B and the bond+gold diversifier (S-039) are out of scope; can be re-blended in a later version.

Process (exploration → freeze → confirm)

1. Exploration (first, free): implement A & B in the exploration workspace, sweep on in-sample ≤2022, measure hedge size, turnover and net@3 Sharpe; iterate. No ID/trial/verdict.
2. Freeze (owner "go"): fix the chosen A-mechanism + B-levers + the Stage gate; then assign the run.
3. Confirm = FORWARD PAPER only — the 2023+ holdout block is spent; no holdout re-runs. Track forward against the frozen gate.

Exploration findings (2026-07-13 — pre-freeze, no verdict, no trial)

Ran A & B on the S-038 base (stocks_only, κ-size, sector-tilt, bands 1.25/0.50, ≤2022) in the exploration workspace (engine diagnostics/levers added: return_hedge, fbal_lambda, rebal_k, sector_balance).

Baseline burdens quantified. ETF hedge (net factor exposure) ≈ 2.45× a leg (median 2.0×, p90 4.4×, max 16×, std 1.6× — large AND variable). Weighted turnover ~23%/day. net@3 Sharpe 0.72, realised beta +0.09.

Component B — lower turnover. The clean lever is lowering s_close (hold to fuller reversion), NOT raising s_open (which drops names, breaks breadth, kills Sharpe — as the spec warned). s_close 0.50→0.25: turnover −21% (23→18%/day), net@3 0.72→0.64, breadth preserved (180→208 names).

Component A — the hedge is stubborn. - Factor-neutral WEIGHTING (daily): kills the hedge (2.45×→~0) but the daily reweight explodes turnover (23→32%) → net@3 0.72→0.53. Rebalancing less often does NOT rescue it — the book churns daily, so held neutrality breaks and the hedge drifts back to ~base. Weighting trades the hedge burden for a turnover burden. - Sector-balanced SELECTION (owner's pick): equalises long/short counts per dominant sector. Does not reduce the hedge (2.45× unchanged — sector counts ≠ weighted factor β; the market factor, within-sector βs and κ-weights are untouched). BUT a free Sharpe boost: net@3 0.72→0.80, same turnover/hedge, 180→143 names (signal-concentrated). A quality filter, not a hedge fix. A weighted-β-targeting selection is untested (deferred).

Pragmatic S-040 (owner decision 2026-07-13): sector-balanced selection + s_close=0.25, hedge accepted.

A Pareto improvement over S-038 in-sample: same/slightly-better net Sharpe (0.73 vs 0.72), −21% turnover, market-neutral (β +0.09). The ~2.3× ETF hedge is retained (accepted). IN-SAMPLE ≤2022.

OOS 2023+ check (owner call 2026-07-13 — a legitimate touch: S-040's specific params are unseen on 2023+)

Ran the frozen configs on the 2023+ block (853 days). This reversed the in-sample recommendation — the whole point of an OOS read:

• Core reversion edge is robust OOS (net@3 > 1.1, β ≈ 0) — it works out-of-sample.
• Sector-balanced selection is OOS-HARMFUL (in-sample 0.72→0.80 was overfit; OOS base 1.28 → sel-only 1.01). Dropped. Component A yields nothing usable (weighting costs turnover; selection hurts OOS).
• B-only (s_close=0.25) is the winner: OOS net@3 1.22 vs base 1.28 (−0.06 only), turnover −21%, breadth preserved (218 names), market-neutral.

Revised S-040 = Component B only (lower-turnover recalibration of S-038)

s_close = 0.25 on the S-038 base; no factor-balancing (A didn't pan out); ~2.3× ETF hedge accepted. −21% turnover for ~−0.06 net@3 Sharpe (OOS), market-neutral. Caveat: 2023+ was a favourable reversion regime (base OOS 1.28 vs in-sample 0.72) and NOT a pristine first-touch (reversion family touched it via D-008/009/010) → deflate the 1.2+ hard; forward paper remains the ultimate confirmation. Still pre-freeze.

Stage gate — [CALIBRATE] (freeze before running)

Not set. Suggested shape only (owner to freeze the values): forward ≥12 mo · net Sharpe ≥ [CALIBRATE] · |beta| ≤ [CALIBRATE] · hedge ≤ [CALIBRATE]× a leg · turnover ≤ [CALIBRATE]/day · sleeve-A realised exec cost ≤ [CALIBRATE] bps/side.

Outcome (owner call 2026-07-13): exploration closed, NOT frozen

Judged a middenmoter — a marginal operational refinement of S-038, not a new edge: - Component A failed — the whole reason to build S-040. Factor-neutral weighting trades the hedge for turnover; sector-balanced selection was in-sample overfit and OOS-harmful. The ~2.3× ETF hedge stays. - Component B works but is cosmetic — s_close=0.25 buys −21% turnover for ~−0.06 net@3 (OOS), without changing the return profile or touching the real burden (the hedge). - Action: no forward-paper spin-off. Fold the s_close=0.25 turnover finding into S-038 as a config note; leave this file as the exploration record. Energy moved to higher-upside work.

Registered pre-freeze; exploration closed un-frozen. Do NOT run, freeze, or fill [CALIBRATE].