How the engine works (shared basis)
*(Descriptive — added 2026-07-20; gate/confirmation unchanged. Editable master figure alongside the PNG:
Picture each stock tied by a rubber band to its "fair value" — the mix of the market (SPY) and its sector that normally explains where it trades.
Sleeve A — the daily rubber band (70%). When a stock stretches too far from its twin, the band tends to pull it back, and we bet on that snap-back. The distance is an s-score: past ±1.25 we open (short the one stretched too high, buy the one stretched too low) and we close once it returns inside ±0.5. We run this across ~188 stocks a day, longs ≈ shorts, each bet tiny, so no single name can hurt the book — and only on names whose band is springy enough to revert within 30 days (the κ-filter). It wins most often when the market is choppy and stocks keep reverting.
Sleeve B — the monthly trend rider (30%). A slower, complementary engine: once a month it ranks the 1,000 most liquid stocks by their ~9-month trend (skipping the last month), buys the strongest and shorts the weakest, and otherwise sits still. Its brake is the risk-off switch — if the market turns hostile (SPY down >10% over a year, or unusually wild) it steps aside and holds nothing. It wins when strong stays strong — a trending market.
The blend (70/30). Because every bet is hedged and longs ≈ shorts, the whole book is market-neutral (β ≈ 0) — it barely cares whether the market rises or falls. And because reversion shines in choppy markets while momentum shines in trending ones, the two together are steadier than either alone, dialled to roughly 10% up-and-down swings a year.
On liquidity (why slippage stays bounded). Both universes are chosen for liquidity, not despite it: Sleeve A trades the top-500 by dollar volume and Sleeve B the top-1,000 by trailing dollar volume (3-month average, $5 price floor, refreshed monthly) — so even the 1,000th name is a mid-cap that turns over real volume, not a micro-cap. Sleeve B trades only the momentum extremes (top/bottom deciles, ≈100 names each) and rebalances just monthly, so its turnover — and its cost — is small. The genuinely cost-sensitive engine is Sleeve A (daily, needs marketable ~3 bps fills; passive limits adverse-select), and it was stress-tested directly: the combined book's holdout Sharpe holds from 1.47 at 3 bps to 1.11 at 10 bps + a 1-day fill lag. The one real-world residual is borrow/spread on Sleeve B's beaten-down short decile — bounded (still top-1,000 liquid names) but the place to watch first.
What's new in this version
Everything is inherited from S-038 unchanged except sleeve A's two tweaks below — same sleeve B, 70/30 blend, risk-off switch, s-thresholds (1.25/0.5), top-500 universe, market-on-close, net@3bps, and the sector-reversion tilt.
1. κ exit filter 30d → 60d — the load-bearing change
S-038 used one κ threshold for both entry and exit (<30 days). The per-trade study showed it ejects winners early: of the ~6,300 "ineligible" exits, 92% revert into ±0.5σ within ~20 days (median 8) after the forced close — the dislocation held; we left too soon. Loosening the exit eligibility to <60d cuts ineligible misses 6,302 → 776 (−88%) and lifts sleeve-A Sharpe 0.45 → 0.50 with the tail unchanged (worst-month, CVaR, max-DD all flat on a vol-scaled comparison). Robust: PBO 0.06 / DSR 0.97. Hysteresis and a min-hold were tested and dropped (no extra Sharpe or tail benefit).
2. +macro residual — a marginal, non-overfit cleanup
A curated macro set (rates / credit / commodity / gold) added to the ETF-factor residual — real but marginal (PBO 0.13, DSR 0.91 standalone), so bundle-only, never a standalone edge. Style factors and a naive lasso hurt and are dropped.
How to read the trade charts (admin trade gallery): each round-trip is drawn as its daily s-score — how far the stock strayed from its ETF-twin, in σ. Red dashed = ±1.25 entry, green band = ±0.5 exit target, filled dot = entry / open ring = exit. The paired panel is the actual price vs its ETF-twin — the gap between those two lines IS the dislocation the s-score measures. A textbook trade dives from ~±1.5 into the green band; the κ<60 fix is visible as the misses reverting just after the old forced exit.
How it earned a forward slot
Combined 70/30 (sleeve A at κ<60 + tilt), vol-scaled to 10%/yr: Sharpe 1.11 · AnnRet +8.9% · maxDD 23% over 19.7 years (2004–2026), deepest drawdown Sep-2020 → Feb-2023 → Oct-2025 (61 months). Split: ≤2022 Sharpe 1.01 / DD 21%; 2023+ Sharpe 1.52 / DD 7% — **shown for information only, NOT a fresh monthly-return histogram are on this card.
Forward record so far
Forward paper-trading against the frozen gate is under way; the record here is machine-generated from the live track and updates on its own. It is interim — no result is claimed until the gate's window completes.
| track | cumulative net return |
|---|---|
| S038 — the parent (κ<30) | +2.30% |
| S055 — this successor (κ<60) | +2.41% |
| difference | +0.12 pp |
Gate clock: 13 counted sessions of the 12 months the frozen gate requires.
Conditions & caveats
- 2023+ is not a fresh confirmation — the config was derived with knowledge that already touched that block, so it is illustration, not proof. The real test is forward paper.
- The macro change is marginal (DSR 0.91) — carried for cleanliness, not claimed as an edge.
- Deflate for the cumulative statistical arbitrage family trial count.
The promotion gate
net Sharpe ≥ 0.6 · |beta to SPY| ≤ 0.15 · maxDD ≤ 20% · both sleeves net-positive · sleeve-A realised execution cost ≤ 4 bps/side · AND it must not underperform the S-038 forward track on the same window (it has to earn its extra complexity, else S-038 stands).
Frozen on owner "freeze & publish" 2026-07-23. A change to the base trade is a new slug. Confirmation: forward paper alongside S-038.
Individual trades — the mechanism in action
A sample of real round-trips drawn as s-score paths + price-vs-ETF-twin. It is also why S-055 loosens the κ exit filter: the “misses” mostly revert just after the old forced exit.
Misses — ineligible (dislocation persisted / squeezed)
entry s -2.07 → exit s -1.22 · 2008-09-25 → 2008-10-20
entry s -1.90 → exit s -1.18 · 2008-06-19 → 2008-07-09
entry s -1.46 → exit s -0.68 · 2022-05-12 → 2022-06-09
entry s +1.31 → exit s +1.10 · 2020-07-21 → 2020-08-10
entry s -2.03 → exit s -1.19 · 2007-11-20 → 2007-12-17
entry s -1.42 → exit s -1.27 · 2006-02-01 → 2006-02-21
entry s -1.48 → exit s -1.51 · 2019-07-17 → 2019-08-13
entry s +1.41 → exit s +nan · 2020-10-26 → 2020-11-13
entry s +1.73 → exit s +1.07 · 2008-09-16 → 2008-10-02
entry s +1.33 → exit s +1.27 · 2016-12-16 → 2017-01-25
Reverted but lost money (longer holds, friction/adverse path)
entry s -1.28 → exit s -0.45 · 2011-09-13 → 2011-10-17
entry s +1.38 → exit s +0.28 · 2008-07-08 → 2008-08-27
entry s -1.45 → exit s +0.08 · 2009-01-23 → 2009-03-12
entry s +1.39 → exit s +0.49 · 2012-10-19 → 2013-01-03
entry s -1.59 → exit s -0.21 · 2022-03-09 → 2022-04-18
entry s -1.27 → exit s -0.19 · 2022-04-01 → 2022-05-25
entry s -1.27 → exit s +0.21 · 2009-09-09 → 2009-11-24
entry s +2.61 → exit s -0.11 · 2009-10-23 → 2009-12-04
entry s +1.50 → exit s +0.47 · 2005-11-04 → 2006-01-26
entry s +1.30 → exit s +0.44 · 2005-08-31 → 2005-10-12
Textbook reversion (entered ~1.5σ, closed inside ±0.5) — for contrast
entry s +2.53 → exit s +0.49 · 2021-11-02 → 2021-11-09
entry s +1.26 → exit s +0.10 · 2008-05-15 → 2008-05-21
entry s +1.56 → exit s -0.47 · 2022-01-31 → 2022-02-02
entry s -1.31 → exit s +0.23 · 2013-01-17 → 2013-01-24
entry s +1.61 → exit s -0.46 · 2020-03-26 → 2020-03-30
entry s +1.39 → exit s +0.23 · 2009-04-03 → 2009-04-13
entry s +1.39 → exit s +0.11 · 2006-08-22 → 2006-09-22
entry s -3.26 → exit s -0.22 · 2020-02-26 → 2020-03-20
entry s -1.40 → exit s +0.20 · 2014-02-19 → 2014-02-25
entry s -1.40 → exit s +0.03 · 2022-12-13 → 2022-12-28
Running it on margin — does leverage give you L× the return?
A fair question for a market-neutral book: if you trade it on a margin account at 2×, 3× or 4× your own money, do you get that multiple of the return and the drawdown? The intuition is half right. The tables and chart below are computed on S-055’s own monthly backtest path (same series as the equity curve above), charging 6.5%/yr financing on the borrowed capital.
| Leverage | “L× return” (the naive guess) | Actual annual return | Worst drawdown | Sharpe | Worst month |
|---|---|---|---|---|---|
| 1× | +9% | +8.9% | -23% | 1.10 | -5% |
| 2× | +18% | +10.5% | -50% | 0.70 | -10% |
| 3× | +27% | +11.4% | -70% | 0.57 | -16% |
| 4× | +36% | +11.6% | -83% | 0.50 | -21% |
Two things break the tidy “4× means 4×” picture:
- The return does not scale up. Compounded return goes from ~9% (1×) to only ~12% at 4× — not ~36%. This is volatility drag: you compound the swings, and bigger swings lose more on the way down than they gain back. The return even peaks near 3.8× and then declines — past that point more leverage means less money and far more risk.
- The drawdown does scale — brutally. The worst drawdown deepens from −23% (1×) to −50% at 2× and −83% at 4×. Past roughly about 2.0× the historical worst drawdown already reaches the margin-call zone (shaded), where the broker force-liquidates you at the bottom — so the “it recovers afterwards” scenario never happens for you.
Assumptions: leverage = multiple of the book’s P&L swings on your equity; 6.5%/yr financing on the borrowed turns (real all-in margin + short-borrow cost is often higher); margin-call zone drawn at a conservative −50% equity drawdown. This uses month-end data, which understates the risk — real margin calls fire intraday on maintenance margin, so liquidation would hit earlier and deeper than shown. Note the book already runs ~2× gross by construction (long + short); the leverage here is on top of that. Illustration on backtested data, not advice — S-055 is a forward-paper candidate, not a live product.
Methodology appendix — gates, exact parameters, look-ahead audit — is visible to subscribers. See the plans →