Market-neutral: idiosyncratic reversion + residual momentum

Two orthogonal, market-neutral edges combined: sleeve A buys stocks that are cheap versus their own sector-ETF replication and shorts the rich ones (idiosyncratic mean-reversion), and sleeve B is residual momentum. Because one reverts and the other trends, the blend diversifies — the aim is a few percent a year uncorrelated to the market.

Growth of $1 — in-sample then sealed holdout (vol-scaled to 10%/yr)

Net of cost, vol-scaled to a common 10%/yr so the shape is comparable. The shaded block (2023+) is the sealed holdout the design never saw. Both sleeves and the blend keep climbing through it — but see the honest caveat below on reading the holdout.

Why it should work — the mechanism

Sleeve A (the stat-arb / “pairs vs the sector ETF”). Each stock is explained by a basket of liquid ETFs (SPY + the 9 sector SPDRs) via a rolling 60-day regression — think of it as the stock’s ETF twin. Subtract the twin’s move and what’s left is the stock’s idiosyncratic residual — the part not explained by the market or its sector. That residual tends to over-shoot and then mean-revert over a few days (short-term over-reaction / liquidity provision). We accumulate the residual into a running series, fit a mean-reverting (Ornstein-Uhlenbeck) process, and score how far it has strayed from its own equilibrium as a dimensionless s-score = (deviation) ÷ (its own volatility). |s| > 1.25 = interesting: the stock is ~1.25+ standard deviations cheap (or rich) versus its ETF twin — a valuation gap wide enough to fade. We only trade names that revert fast enough (mean-reversion time under 30 days) and size the strongest, fastest reverters larger, tilted toward the sectors that revert most (health/utilities/materials).

Sleeve B (residual momentum). The mirror image on a longer horizon: rank stocks by their 9-month beta-adjusted return and go long winners / short losers, with a risk-off switch that sits out drawdown/high-vol months. Crucially it is orthogonal to sleeve A (in-sample correlation +0.12): one fades short-term over-reaction, the other rides longer trends. Combining them lifts the risk-adjusted return above either alone.

What one trade looks like

Suppose XYZ (a healthcare name) sells off 6% on no company news while its sector twin (mostly XLV) is flat. Its residual drops; the accumulated residual is now s = −2.1 — cheap vs its twin. The book buys XYZ and shorts the beta-weighted ETF basket that replicates it, so the position is market- and sector-neutral: it only wins if XYZ closes the gap to its twin, not on market direction. We enter market-on-close (the reversion happens overnight, so we must be in by the close). We exit — take profit — when s reverts back inside −0.5 (the gap has closed), also market-on-close; typical hold ≈ a week. If instead XYZ’s twin richens to s = +2.1 we do the reverse (short XYZ, long the basket).

Stop-loss & profit-take

The profit-take is the exit rule itself: close when the residual reverts to |s| < 0.5 — you took the gap. We tested it earlier: closing earlier (|s| < 0.75) captures a touch more of the front-loaded move; waiting for full reversion (|s| → 0) gives it back — so the mid exit is the sweet spot. On stop-loss: an explicit “bail if it diverges to |s| > 3–4” or a fixed %-loss stop did not help in testing — the mean-reversion exit already closes losers, and a “broken pair” (a stock that structurally re-prices and never reverts) is handled by the κ-filter (drop names whose reversion is too slow) and by breadth: ~180 tiny positions mean no single broken pair can sink the book. Sizing is not equal-weight but scaled by reversion speed and sector.

Backtest — in-sample vs the sealed holdout

Read honestly. The holdout (40 months) beats in-sample — treat that as a favourable window, not the true level: deflate the 1.50 toward the in-sample ~0.9-1.0. It is a forward-paper candidate confirmed once on unseen data (market-neutral, both sleeves positive), not a proven live performer. Sleeve A’s edge is real gross but cost-sensitive (18%/day turnover; net Sharpe ≈ 0.72 at 3 bps, 0.28 at 5 bps) — which is why execution is market-on-close and the forward paper measures the realised cost against a ≤4 bps/side gate.

Borrowing cost

The short leg needs stock borrow, but sleeve A shorts liquid top-500 names (overbought vs their twin), which are almost all easy-to-borrow at general-collateral rates (~0.3–1%/yr) — a small drag, unlike hard-to-borrow small-caps. It is not yet inside the headline net figure (which charges the 3 bps spread); the forward paper measures the real borrow so the Stage-3 economics are honest. Expected impact: on the order of a few tenths of a percent per year.

Deploying it — return stacking & capital efficiency

Because the book is market-neutral (beta ≈ 0), it is an alpha you stack on top of other exposure rather than instead of it — the return-stacking / portable-alpha idea. You keep your normal base (equity beta, or T-bills for carry) and run this neutral book on the same capital, so one dollar earns two return streams: the base + the uncorrelated alpha. A neutral strategy is the ideal thing to overlay precisely because it adds no market direction.

Why the capital efficiency works. A dollar-neutral long-short is efficient with margin: shorting generates cash proceeds that partly finance the long side, so net financing is low. Under portfolio margin (which nets the tiny net exposure) a modest capital base can carry a meaningful gross book — the classic “$100 supports ~$200 long / ~$200 short”.

Stated honestly — this is risk-sizing, not free alpha. Leverage scales return and risk equally: a Sharpe-~0.9 book levered 2× is still Sharpe ~0.9 but with ~2× the volatility and drawdown (and margin-call risk). The full “$200/$200 on $100” efficiency needs portfolio-margin approval — standard Reg-T is heavier. Short borrow and financing are small but not zero. Size to a drawdown you can actually sit through; the stack/leverage is capital efficiency, not a higher Sharpe.

Optional: bond + gold diversifier (S-039)

A companion setup S-039 stacks a directional bond+gold sleeve (equal-weight IEF + GLD, 30%) on top of the neutral book. It is not a pairs trade — it holds bonds and gold outright, uncorrelated assets that cushion equity stress. The book stays equity-neutral (beta to SPY −0.12, within the neutral gate) but now carries some duration and gold-price risk — which is the point (a market-neutral spread can’t hedge a drawdown). Confirmed on the sealed holdout: in-sample Sharpe 0.98 → 1.09, holdout 1.50 → 1.83 (deflate heavily — a gold-favourable window; bonds themselves crashed −31% in 2022).

Both lines vol-scaled to 10%/yr, so the higher-Sharpe blend (teal) climbs above the neutral book (black) at the same risk; both stay ≈beta-0 through SPY’s corrections (red bands). A higher Sharpe is spent as either more return or less drawdown — a sizing choice, not both at once.

The universe — stocks & their ETF baskets

The ~500 most-liquid US common stocks, each hedged against its best-fit combination of SPY + the 9 sector SPDRs (its “ETF twin” from the regression). The bars show how many stocks load most on each sector basket, with example tickers and the reversion-tilt weight (darker = tilted harder). Data-derived sector = the strongest ETF loading, used only for the tilt and this chart, not for the hedge; a few communication-services mega-caps (no dedicated SPDR pre-2018) land on their nearest basket.

v2 correction (transparent). The first freeze (S-037) accidentally included ~24% ETFs (bond/gold/EM) in the “stock” universe. This S-038 corrects it to common stocks only — which improved the result (in-sample Sharpe 0.92 → 0.98, holdout beta −0.07 → −0.03), re-confirmed on the same sealed holdout. The flawed S-037 stays in the public record.

Why it is paper, not live: sleeve A is breadth-dependent (~180 names, daily, market-on-close) so it is automation-gated; the lighter monthly momentum sleeve leads the live path. Lineage: discovery D-009 → this setup. Frictions conservative; positions are private (admin).