What Big Banks & Hedge Funds Actually Trade — and What Our Engine Can Replicate

Date: 2026-10-21 · Scope: replicability audit for a long-only, daily-bar (OHLCV) research study over S&P 500 ∪ Nasdaq-100 (~518 names), with earnings dates and a basic fundamentals snapshot. No tick data, no historical options chains, no co-location, no short book.

Bottom line up front: The durable institutional edge is not a secret indicator. Bank money is made on market structure (spreads, execution, financing, balance sheet) — none of which we can touch. Hedge-fund/CTA money, at the documented core, is made on a short list of boring, published factor premia and behavioral effects — cross-sectional momentum, trend following, value, quality/profitability, post-earnings drift — run with industrial risk control, diversification, and leverage. Roughly half of that core is replicable with our data. The other half (market making, stat-arb pairs at scale, merger arb, Medallion-style short-horizon stat arb) is not, and the honest reasons are structural, not a lack of cleverness.

The four strategies recommended for the engine, ranked: (1) cross-sectional 12-1 momentum sleeve, (2) PEAD earnings-drift sleeve, (3) time-series momentum with volatility targeting, (4) quality+value composite (tracked forward only until point-in-time data accrues). Exact rule sets in §5.

0. Our baseline — what the engine already is

So that "fit" and "diversification" judgments below are grounded, the current state (from ~/workspace/ideas-engine/):

• Universe/data: ~518 current S&P 500 ∪ Nasdaq-100 constituents, daily OHLCV (yfinance), 3y price cache (prices_3y.parquet), earnings dates (Nasdaq calendar + cache), fundamentals = a current snapshot per ticker (trailing/forward P/E, EV/EBITDA, debt/equity, revenue growth, gross/operating/profit margins, market cap) for all 518 names.
• Study mechanics: long-only setups tracked as research — each setup scored against ATR-based stop/target levels, per-strategy max-hold time stops, an earnings blackout (no setup counted within 5 sessions of earnings), and SPY-regime gating (SPY < 200DMA zeroes "chop" strategies).
• Existing strategies: breakout 20d-high (benched, −16.8% over 2y backtest), mean-reversion RSI(14)<40 (+14.4%), golden cross (+12.2%), MACD bull-cross (+15.1%), novel signals rsl (20-day relative-strength leader pullback), uva (up-volume accumulation), vcve (volatility-compression breakout; observe-only until 30 backtest trades).
• Learning machinery already present: learn_3y.py (recency-weighted profit-factor weights, capped at 2.0), walkforward.py (train 2023–25 → test 2025–26 out-of-sample), regime multipliers, a killed-strategy precedent (gh, zeroed 2026-10-15).

Two structural facts matter for everything below: (a) every existing strategy is a price/volume technical trigger — the strategy set has zero fundamental and zero event-driven exposure; (b) the fundamentals cache is a snapshot, so it can drive live signals honestly but cannot backtest value/quality without look-ahead bias; and the universe is current constituents, so all backtests carry survivorship bias (the walk-forward script already flags this).

1. Bank desks — mostly not replicable, and why

1.1 Market making — NOT replicable

What it is: continuous two-sided quotes in equities/options/futures; profit = bid-ask spread × turnover − adverse-selection losses − inventory risk. Run by Citadel Securities, Virtu, Jane Street, and bank equities desks. The inputs are queue position in the order book, sub-millisecond latency, co-location, maker rebates, and inventory models of the Avellaneda–Stoikov / Glosten–Milgrom family.

Why we can't: every one of those inputs is absent. With daily bars we never see the spread, the book, or the queue — the entire spread profit is invisible at our resolution. This is also where the "secret algorithm" mythology is most wrong: Renaissance's Medallion fund — the usual Exhibit A — earned ≈66%/yr gross, ≈39%/yr net (1988–2018, as reported in Zuckerman, The Man Who Solved the Market, 2019) from thousands of tiny, short-horizon statistical edges (holding periods of minutes to days) on proprietary cleaned tick history, with near-perfect execution and substantial leverage, in a fund closed to outsiders since 1993 and capped around $10B. It is a data-and-execution business, not a formula. Virtu's 2014 IPO filing famously reported one losing trading day in 1,238 — that is what spread capture at scale looks like, and it is a latency/infrastructure statistic, not a signal statistic.

1.2 Execution algorithms (VWAP / TWAP / POV / implementation-shortfall) — replicable as discipline, useless as alpha

What they are: order-slicing schedules that minimize market impact and slippage versus a benchmark (arrival price, day VWAP). Theory: Perold's implementation shortfall (1988); Almgren–Chriss optimal execution trajectories (2000). Every bank algo desk (and every large fund's trading desk) runs these.

Verdict: they reduce cost; they create no return. At our scale (daily bars, prices already assumed at close ×1.001) there is nothing to optimize — but there is one transferable practice: measure and charge ourselves realistic execution cost. Our backtest's near-zero cost assumption flatters every strategy and is disqualifying for high-turnover ideas (see short-term reversal, §2.5). Recommendation carried into §4: add a flat per-side cost haircut (e.g., 5–10bp for large caps) in the simulator before trusting any new strategy's numbers.

1.3 Index / ETF arbitrage — NOT replicable

Futures-vs-cash basis trades and ETF create/redeem arbitrage are basis-point edges captured in seconds, requiring futures access, authorized-participant status or prime-brokerage financing, and intraday data. The one daily-scale residue — index reconstitution (announcement effects for S&P 500 additions/deletions) — was a real documented trade in the 1990s but has been heavily arbitraged down since; treat as trivia, not a program.

1.4 Factor risk models (Barra-style) — REPLICABLE, and the most valuable bank technology we can steal

Barra (MSCI) models decompose every stock's return into factor exposures (market beta, size, value, momentum, volatility, leverage, liquidity…) estimated by cross-sectional regression, producing a covariance matrix used for risk forecasting, hedging, and portfolio construction. Banks and funds use them to control risk, not to find alpha directly.

A simplified version is fully feasible with our data panel: monthly cross-sectional regressions of returns on standardized characteristics (beta, log market cap, trailing P/E, 12-1 momentum, 60d realized vol — all available). Concrete uses for our book: - Concentration alarms: flag when all 10 tracked slots are effectively the same factor bet (e.g., high-beta momentum) — a real failure mode of the current trigger-based strategy set. - Risk attribution in the daily scoring: report how much of the tracked result is market beta vs. idiosyncratic. - Cleaner strategy evaluation: a signal that only "works" by loading on beta in a bull sample should not earn weight.

This is infrastructure, not a strategy — but it is genuinely what separates institutional books from retail ones.

(One-line footnote: post-Volcker, banks barely run directional prop books at all. "Bank trading algorithms" in 2026 ≈ market making + execution + risk models. The directional money sits in hedge funds and CTAs — §2.)

2. Hedge funds & CTAs — the documented core, strategy by strategy

2.1 Statistical arbitrage & pairs trading — real, decayed, structurally awkward for us

The canonical method (Gatev, Goetzmann & Rouwenhorst 2006): 12-month formation period; normalize each stock's price path; for every pair compute the sum of squared deviations between normalized paths; take the 20 closest pairs; trade each pair for 6 months (overlapping portfolios): open when the pair's prices diverge by >2 standard deviations of the historical spread, close on convergence (or at period end). The cointegration variant (Vidyamurthy 2004; Rad, Low & Faff 2016) selects pairs whose spread is statistically stationary (Engle–Granger/Johansen) and trades z-score thresholds on the hedge-ratio-adjusted spread.

Evidence: GGR documented excess returns up to ~11%/yr (1962–2002 sample). Then the decay literature is unusually unanimous: - Do & Faff (2010): profitability declined steadily; after transaction costs the simple distance method loses money from ~2002 onward; the cause is less reliable convergence (pairs "divorce"), not just crowding. - A full 1962–2020 replication of GGR (RStudio/CRSP) concludes the simple strategy's profitability "has essentially disappeared." - Cointegration-based selection degrades more gracefully (Rad, Low & Faff 2016; Clegg & Krauss 2018 partial cointegration) — better pair quality, fewer non-convergences — but still a fraction of the 1990s economics.

Failure modes: non-convergence/structural breaks (one leg's fundamentals change permanently), crisis correlation spikes (spreads blow out exactly when capital is scarce), borrow costs and recall risk on the short leg, crowding.

Fit for us: poor. It needs a short leg (our book is long-only), a dedicated pair book (our 10 slots are shared), and its best-documented version is dead. A "pairs-lite" (long only the laggard leg of a cointegrated pair) discards half the construction and most of the evidence. Verdict: do not build now. If it is ever revisited, cointegration selection or nothing.

2.2 Cross-sectional factor investing — the AQR / Fama-French core

This is the largest, most replicated body of evidence in all of quantitative finance, and the backbone of AQR, DFA, Bridgewater-style equity books, and most multi-manager platforms (in long-short, factor-neutral form).

(a) Momentum 12-1. Jegadeesh & Titman (1993): rank stocks on trailing 6–12 month returns, buy winners/sell losers; the original sample earned ~1%/month gross long-short, robust across formation/holding combos. The modern standard skips the most recent month (12-1) to avoid the short-term reversal of §2.5. Asness, Frazzini, Israel & Moskowitz ("Fact, Fiction, and Momentum Investing," 2014) put the 1991–2013 momentum premium at ~6.3%/yr (vs ~3.6% for value) and document it across ~40 countries and multiple asset classes over two centuries of data. Failure mode — momentum crashes: at sharp market reversals, the loser leg (beaten-down, high-beta names) rips upward; the momentum factor lost roughly 40%+ in 2009 (Daniel & Moskowitz, "Momentum Crashes," JFE 2016). Documented mitigations: volatility-scaling the momentum exposure (Barroso & Santa-Clara 2015 — risk-managed momentum roughly doubled the Sharpe in their sample, figure approximate), and combining with value. Decay: momentum has survived publication better than most anomalies, but McLean & Pontiff (2016) is the standing warning for all published predictors: returns decay ~26% out-of-sample and ~58% post-publication on average across 97 predictors. Plan on capturing a fraction of the printed number. Long-only haircut (important): the literature numbers are long-short. A long-only top-quintile book keeps only the winner leg — historically the steadier but smaller half of the premium, plus full market beta. Expect "market + a modest tilt," not 6.3% of alpha.

(b) Value. Fama–French HML (book-to-market; earnings/price variants similar): ~3.6%/yr premium 1991–2013 per the AQR summary above; a century of support but with brutal decades (2010s US large-cap value drought). Needs point-in-time fundamentals — see the data warning in §0/§5.4.

(c) Quality / profitability. Novy-Marx (2013): gross profitability (gross profits/assets) predicts returns as powerfully as value — top-minus-bottom spread ≈0.31%/month, Fama-French-3 alpha ≈0.52%/month (t ≈ 4.5). AQR's Quality-Minus-Junk (Asness, Frazzini & Pedersen 2019) generalizes it (profitability, growth, safety, payout). Two properties make quality precious for a portfolio: its signal is nearly uncorrelated with price-based signals (Novy-Marx notes the profitability signal is negatively correlated with valuation signals — value and quality hold very different stocks and offset each other), and it tends to hold up in drawdowns ("flight to quality").

(d) Low volatility / betting against beta. Frazzini & Pedersen (2014): leverage-constrained investors overpay for high-beta stocks, flattening the security market line. In their data (1926–2012) the lowest-beta decile earned a Sharpe of ~0.70 vs ~0.28 for the highest-beta decile — similar returns, a third of the risk. As a long-only strategy this is a defensive sleeve (market-like returns, lower vol), not an alpha engine; the full BAB factor needs leverage on the low-beta side and a short high-beta side. Verdict: not a top-4 alpha candidate; legitimate as a Tier-1 ballast concept later.

2.3 Time-series momentum / managed-futures trend — the CTA core (Man AHL et al.)

Method (Moskowitz, Ooi & Pedersen 2012): for each instrument, go long if its own trailing 12-month excess return is positive, short if negative; size each position inversely to its volatility (volatility targeting) so every instrument contributes equal risk. Evidence: MOP's diversified portfolio across 58 futures markets (1985–2009) delivered Sharpe ≈ 1.0 gross of costs (as reported; treat as approximate). Hurst, Ooi & Pedersen ("A Century of Evidence on Trend-Following Investing," 2017) extend to 67 markets, 1880–2016: Sharpe 0.76 after estimated costs and 2-and-20 fees, positive net returns in essentially every decade, and positive performance in 8 of the 10 worst equity drawdowns — the famous "crisis alpha." Average per-market Sharpe ≈ 0.4.

Honest counterweights (must travel with the headline): - Kim, Tse & Wald (2016) and later replications find TSMOM much stronger pre-2001 and weak in equities post-2009; one MOP-methodology replication found Sharpe 0.75 vs 0.73 for passive over its sample with negative alpha for 2009–2021. - A well-known deconstruction (QuantPedia summary of the literature): much of MOP's magnitude comes from the volatility scaling itself — unscaled, the average alpha falls from ~1.3%/mo to ~0.4%/mo. The signal is real but thinner than the headline; the risk management is load-bearing. - The original evidence is futures across four asset classes. Single-asset-class, long-only equity trend is the weakest version of it — no shorts, no cross-asset diversification. In our book its job is drawdown diversification, not lead returns.

Failure modes: whipsaw in range-bound markets (the 2010s CTA winter), V-shaped reversals (long/flat exits late, re-enters later), vol-target procyclicality.

2.4 Post-earnings announcement drift (PEAD) — the event-driven basic that fits us best

Phenomenon: stocks continue drifting in the direction of an earnings surprise for weeks after the announcement (Ball & Brown 1968 — the oldest documented anomaly). Bernard & Thomas (1989): sorting on standardized unexpected earnings (SUE), the top-minus-bottom decile spread earned ~4.2% over the 60 trading days post-announcement, positive in 41 of 48 quarters (1974–85), including down-market quarters; drift is monotonic in the surprise and ~25–30% of it clusters around the next announcements. Fama (1998) called it the "granddaddy of underreaction events."

Decay, honestly: PEAD has shrunk, is weakest exactly in large liquid names (subsequent literature, e.g., Chordia et al., finds the drift in big, heavily arbitraged stocks much reduced), and it disappeared/inverted around the Great Recession (Anderson & Ivanov 2019). The small-cap literature numbers (8–9%/quarter hedge returns in 1990s samples) do not transfer to S&P 500 names. What remains: a modest, well-understood drift in the most-watched universe on earth — worth harvesting at a tracked-study scale precisely because our engine already knows every earnings date (it runs an earnings blackout today; PEAD is the mirror image: enter after the print instead of avoiding it).

Failure modes: guidance cuts overwhelming a headline EPS beat; "beat and fade" in high-multiple names; surprise measurement error (which consensus? as-reported vs adjusted EPS); overnight gap risk is mostly before entry in our construction (we enter after the announcement reaction).

2.5 Short-term reversal — real, but it's a liquidity-provision wage, not a free anomaly

Jegadeesh (1990) / Lehmann (1990): last month's (or week's) losers beat winners by ~2%/month gross in early samples. The follow-through literature is brutal about capturability: most of the gross profit sits inside the bid-ask spread (Kaul & Nimalendran; Ball, Kothari & Wasley) and Avramov, Chordia & Goyal (2006) show costs kill it in the small, volatile names the sort selects. The constructive version — de Groot, Huij & Zhou (2012): restrict to large caps and optimize turnover, and reversal still earned ~30–50bp/week net in their sample. Fit for us: weak. Our meanrev RSI<40 already harvests a slower, cheaper cousin of the same overreaction effect; a dedicated 1-week reversal sleeve would multiply turnover against a cost model our simulator barely charges. Verdict: watchlist only — revisit only if the cost model (§1.2) and a large-cap-only, turnover-capped design are in place.

2.6 Event-driven basics — mostly out of reach

Merger arbitrage (buy target below deal price) is a short-volatility insurance premium with binary deal-break risk; it needs deal terms, financing/antitrust assessment, and a short acquirer leg — not buildable from our data, and its "steady small gains, occasional large loss" profile is the opposite of what the two-tier book wants from new strategies. Index additions (§1.3) are arbitraged down. Buyback/spinoff announcements have real literature support but need an events feed we don't have. PEAD (§2.4) is the event strategy our data actually supports.

3. Recursive learning, done properly — the part that decides whether any of this survives contact with reality

The mandate — "recursively learn and correct" — is exactly where retail systems die: they learn in-sample, promote noise, and call it adaptation. The institutional standard (López de Prado, Advances in Financial Machine Learning, 2018, and the Bailey/López de Prado studies) in summary form:

1. Walk-forward is the minimum, not the gold standard. Train on a window, freeze, test on the next unseen window, roll. Our walkforward.py already does one honest split (train 2023–25 → test 2025–26). Extend to multiple rolling folds and report the distribution of out-of-sample results, not the best fold.
2. Count your trials. Every strategy variant and parameter tweak ever tested is a draw from the multiple-testing urn. Harvey, Liu & Zhu (2016), cataloguing 316 published factors, argue a new factor needs t ≥ 3.0, not 2.0 — and that a third to a half of published factors are likely false. A trials ledger (a simple append-only log of everything backtested, win or lose) is the cheapest high-value discipline we can add.
3. Deflated Sharpe Ratio (Bailey & López de Prado 2014). Adjusts an observed Sharpe for the number of trials, sample length, and non-normality, returning a probability the true Sharpe > 0. Promotion gate: DSR ≥ 0.95. Companion: Probabilistic Sharpe Ratio / Minimum Track Record Length (Bailey & López de Prado 2012) — at realistic Sharpes (0.5–0.8), the track record needed for statistical confidence is measured in years. This is the honest answer to "how long until we know": longer than anyone wants.
4. Probability of Backtest Overfitting via CSCV (Bailey, Borwein, López de Prado & Zhu 2017). Combinatorially symmetric cross-validation estimates how often the in-sample-best variant lands below median out-of-sample. Use it whenever we select among variants (e.g., momentum lookback 6 vs 9 vs 12); practitioner convention treats PBO < ~0.2 as acceptable (label: convention, not theorem).
5. Purging and embargo (López de Prado). When labels/holding periods overlap in time (all of our multi-week holds do), training and test samples leak. Purge overlapping observations and embargo a buffer around test windows. For our event trades (PEAD), purge overlapping drift windows around earnings clusters.
6. Assume publication decay. McLean & Pontiff (2016): ~26% out-of-sample, ~58% post-publication average decay. Standing rule: haircut any literature return by ~half when setting expectations, and judge strategies against the haircut, not the headline.
7. Regime conditioning — measured, not vibes. We already gate on SPY vs 200DMA. Upgrade: log every strategy's results by regime bucket (trend/chop/down × vol tercile) so gates are set by measured conditional performance. The literature's one regime result worth hard-coding: momentum/trend crash risk concentrates in high-volatility rebounds after drawdowns (Daniel & Moskowitz 2016) — which is why vol-scaling, not binary on/off, is the documented remedy.
8. Weight-update discipline (formalize what Spring Loader already does informally): update at most monthly; require a minimum sample before any move (≥10 scored trades to move at all — the existing Spring Loader rule; ≥30 to reach full weight — the existing vcve gate); cap steps (≤ ±0.5, as already implemented there); shrink toward neutral (weight = n/(n+k) × measured edge, k ≈ 20); never let an in-sample backtest alone set a weight above 1.0; demote automatically when live out-of-sample profit factor < 1.0 over ≥30 trades (the gh kill, made a rule instead of an event).

Net assessment of our current learning loop: structurally sound skeleton (walk-forward exists, min-sample gates exist, step caps exist in Spring Loader), three gaps: no trials ledger (so no DSR/PBO is even computable), no cost model in the simulator, and weights still partly informed by in-sample backtests (learn_3y computes weights on the same data its picks come from — its own docstring admits this; the walk-forward file exists precisely to correct it). Fix those three before adding any new strategy's weight.

4. What is marketing myth vs documented — one-paragraph honesty box

Myth: "Hedge funds use secret AI/indicators." Documented: the persistent core is published factor premia and trend, run with leverage, breadth, and risk discipline; ML (e.g., meta-labeling at multi-manager platforms) is a sizing/filter layer on top of conventional signals, trained on data breadth (thousands of instruments, decades of clean tick history) that a 518-name daily-bar book does not have. Myth: "Bank algos predict direction." Documented: bank algos mostly execute and make markets; direction is not the product. Myth: published anomaly returns are capturable as printed. Documented: they are long-short, gross of costs, often small-cap-tilted, and decay ~58% after publication on average. Myth: a great 2–3y backtest Sharpe proves an edge. Documented (Bailey & López de Prado): with enough trials, a beautiful backtest is the expected outcome of pure noise; that is what DSR/PBO exist to quantify.

5. Ranking — top 4 for our engine

Scored 1–5 per criterion (5 = best). "Diversification" is measured against the existing book (breakout / RSI mean-reversion / golden cross / MACD / rsl / uva / vcve — all price-volume triggers).

*Quality+value scores 2 on data fit today because the fundamentals cache is a non-point-in-time snapshot: fine for forward-tracked signals, inadmissible for backtesting. It ranks #4 despite top-tier evidence and the best diversification in the set. PEAD edges TSMOM at #2 on diversification (zero trigger overlap) despite slightly weaker modern evidence.

5.1 — #1: Cross-sectional momentum, 12-1, long-only sleeve ("XSMOM")

Rule set (engine-implementable): - Universe: the 518 names with ≥252 sessions of history. - Signal, monthly (last session of month): mom12_1 = Close[t−21] / Close[t−252] − 1 (skip the most recent 21 sessions — the standard Jegadeesh–Titman skip-month). - Selection: rank cross-sectionally; candidates = top quintile (~100 names). From those, fill open slots (at most 10 names) by rank, subject to: price > 200DMA (absolute-momentum guard, Antonacci dual-momentum style), earnings blackout (existing 5-session rule), not already held by this sleeve. - Sizing: engine risk standard, scaled by relative volatility: size ∝ (σ_target / σ_i,60d), capped by the existing 20% position cap. Portfolio crash guard: if SPY 20d realized vol > 1.5× its trailing 1y median, halve new-position sizes (simplified Barroso–Santa-Clara vol management). - Exits: at each monthly rebalance, exit any holding that has fallen out of the top quintile or below its 200DMA. Intra-month: hard stop −15% from entry only (crash brake). Implement with stop = −15%, no fixed target, max_hold = 21 sessions to force the monthly review. - Evidence to expect, honestly: long-short premium ~6.3%/yr (AQR, 1991–2013) is not our number; long-only large-cap captures the winner leg plus beta. Success criterion for the study: beat an equal-weight universe benchmark over rolling 12m with lower drawdown. - Failure modes: reversal crashes (2009-style), sector crowding (all top-quintile = one theme), turnover at rebalance boundaries (mitigated by top-quintile exit hysteresis: keep until out of top quintile, not top-10). - Diversification note: same premium family as rsl, but rsl is a 20-day leader-pullback timing signal with 20-session holds; XSMOM is a 12-month ranking with monthly holds. Expected trade overlap is low; factor overlap is real — count them as cousins, not strangers, in book risk.

5.2 — #2: PEAD earnings-drift sleeve ("PEAD")

Rule set: - Data: earnings date + consensus EPS estimate vs reported EPS (yfinance earnings-dates history where available; cache the values at announcement time — never backfill restated history). Fallback if estimate data is missing for a name: skip the name (do not substitute a price-only proxy silently; a price-reaction variant is a different strategy and must be trialed separately). - Trigger: on the first session after the announcement (enter at that session's close, consistent with engine fill convention), require all of: - EPS surprise (actual − estimate) / |estimate| ≥ +10%; - announcement-day return minus SPY return ≥ 0 (the market confirmed the print — filters guidance-cut fades); - close > 50DMA (drift, not falling-knife catching). - Exits: max_hold 20 sessions (inside the ~60-day drift window; the liquid large-cap compromise — flagged as a deliberate trade-off, since most documented drift for large firms accrues in the first ~60 days); stop = entry − 2×ATR(14); target = entry + 3×ATR(14) (mirrors rsl plumbing so the simulator/ledger need no new machinery). - Weighting: reduced weight until ≥30 scored observations, per house discipline. - Evidence: Bernard & Thomas (1989) spread ~4.2%/60 trading days (1974–85, all-cap); expect a small fraction of that in modern mega-caps — the forward study is precisely to measure what’s left. Earnings season gives ~4 waves/yr × dozens of qualifying names: setup frequency is adequate for the 30-observation gate within ~2–3 quarters. - Failure modes: beat-and-fade on guidance, surprise-data errors, drift arbitraged away in the most liquid names (the null result is a legitimate, cheap answer).

5.3 — #3: Time-series momentum with volatility targeting ("TSMOM sleeve")

Rule set: - Signal, weekly (Friday close), per name: score = sign(r21) + sign(r63) + sign(r252) (1/3/12-month blend, Hurst–Ooi–Pedersen style — blending reduces whipsaw vs a single horizon). Long state: score ≥ +2. Flat: score ≤ 0. (Long-only adaptation of MOP; no shorts.) - Entry/exit: enter at next session's close when a name flips into long state and a slot is free (priority by score, then by 12m return); exit at the weekly check when score ≤ 0. Catastrophe brake intra-week: close below entry − 2.5×ATR(14). - Sizing: volatility-targeted: position $ = slot $ × (12% / σ_name,60d annualized), floored/capped by existing caps; no fixed profit target. - Relationship to golden cross — stated plainly: this harvests the same underlying trend premium as the golden-cross strategy with different mechanics (state-based holding vs event trigger, vol-scaled sizing vs fixed). Run them head-to-head in the walk-forward harness; if TSMOM dominates out-of-sample, retire golden cross rather than double-counting one premium as two strategies. - Evidence: MOP (2012) diversified Sharpe ≈1.0 gross (approximate, futures, long-short); Hurst et al. (2017) 0.76 net of costs+2/20 over 137 years, positive in 8 of 10 worst equity drawdowns. Discount hard for equity-only, long-only: this sleeve's job in our book is crisis-period diversification of a book that currently goes 100% long dips and breakouts into selloffs. - Failure modes: sideways-market whipsaw (expect long flat stretches — the CTA winter is the base case, not the tail), late exits in V-reversals.

5.4 — #4: Quality + value composite ("QV sleeve") — tracked forward only

Rule set: - Data (with the integrity rule): at each quarterly rebalance, use the fundamentals snapshot as of that date (gross margin, operating margin, forward P/E, EV/EBITDA, debt/equity, revenue growth). Starting now, snapshot the fundamentals cache monthly and archive it — within a year we own a genuine point-in-time dataset. Backtesting this strategy on today's snapshot is forbidden (look-ahead); it admits on forward tracked evidence only, exactly like the vcve observe-gate. - Score (cross-sectional z-scores across the 518): quality_z = mean(z(grossMargins), z(operatingMargins)); value_z = mean(−z(forwardPE), −z(EV/EBITDA)); composite = 0.5·quality_z + 0.5·value_z. Guards: exclude worst-quintile debt/equity; exclude revenue growth < −5%. (Proxy honesty: the literature's quality measure is gross profits/assets (Novy-Marx); our cache lacks total assets, so margins are the stand-in. Label it a proxy in the code, as with the engine's other approximations.) - Portfolio: top 10 by composite; quarterly rebalance; exit at rebalance if out of top 20; max_hold 63 sessions; wide stop = entry − 3×ATR(14) (this is a slow factor sleeve, not a trade). - Evidence: Novy-Marx (2013) profitability spread ~0.31%/mo, FF3 alpha ~0.52%/mo (t≈4.5); value ~3.6%/yr (AQR 1991–2013); the two sleeves' signals are near-uncorrelated (~0.1) and hedge each other (Novy-Marx). Long-only, large-cap capture is partial; the realistic prize is a steadier Tier-1 compounder with the lowest correlation to everything else we run. - Failure modes: value traps the quality screen misses, growth-at-any-price regimes (2020-style) where the sleeve lags for quarters, snapshot data errors from the vendor (spot-check against EDGAR filings — free, as-reported, point-in-time — for the top-10 names each rebalance).

Not ranked, with reasons (one line each)

• Pairs/stat-arb: short leg + documented post-2002 death of the simple version; revisit only with cointegration selection and a short-capable book.
• Short-term reversal: largely a bid-ask/liquidity wage; our simulator's cost model can't yet judge it honestly; meanrev already covers the family.
• Low-vol/BAB long-only: a ballast, not an alpha source; candidate Tier-1 risk overlay later.
• Merger arb / index-add / buyback events: data we don't have; index-add effect arbitraged down.

6. Suggested build order (for the parent agent's planning)

1. Cost model + trials ledger (infrastructure first; both are prerequisites for believing any result below).
2. XSMOM sleeve — backtestable today on the 3y cache; run through the existing walk-forward harness before it is tracked forward.
3. PEAD sleeve — backtest is indicative only (thin surprise history); track forward at reduced weight with a 30-observation gate.
4. TSMOM sleeve — backtest today; run as golden-cross challenger in walk-forward.
5. QV sleeve — start monthly fundamentals snapshots immediately; signals tracked forward only.
6. Barra-lite factor exposure report on the tracked setups (monthly) once 1–2 sleeves are running.

Sources

• Moskowitz, Ooi & Pedersen (2012), "Time Series Momentum," JFE — summary & critiques: http://quantpedia.com/strategies/time-series-momentum-effect ; deconstruction: https://quantpedia.com/deconstructing-the-time-series-momentum-strategy/
• Hurst, Ooi & Pedersen (2017), "A Century of Evidence on Trend-Following Investing," JPM (AQR): https://www.aqr.com/-/media/AQR/Documents/Insights/Journal-Article/AQR-JPM-Fall-2017.pdf — Sharpe 0.76 net of costs and 2/20 fees; per-market Sharpe ≈0.4.
• Asness et al., momentum premium ~6.3%/yr vs value ~3.6%/yr (1991–2013), and Daniel & Moskowitz (2016) momentum-crash documentation: https://freedium-mirror.cfd/https://medium.com/@navnoorbawa/aqrs-momentum-factor-crashed-42-in-2009-cfm-s-category-crashed-too-d20bba1def54 ; https://fooletfs.com/insights/three-charts-that-tell-the-momentum-story
• Gatev, Goetzmann & Rouwenhorst (2006) and Do & Faff decline/cost findings: https://QUANTPEDIA.com/strategies/pairs-trading-with-stocks ; replication 1962–2020: https://rstudio-pubs-static.s3.amazonaws.com/762156_9893fc1914ee4feea863872bf65e02bb.html
• Bernard & Thomas (1989/1990) PEAD magnitudes and structure: https://www.eventstudytools.com/earnings-announcements ; https://en.wikipedia.org/wiki/Post%E2%80%93earnings-announcement_drift
• Short-term reversal costs (Avramov–Chordia–Goyal; de Groot, Huij & Zhou large-cap net results): https://repub.eur.nl/pub/25718/AnotherLook_2011.pdf
• Moreira & Muir (2017), "Volatility-Managed Portfolios," Journal of Finance: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2773438
• Novy-Marx gross profitability (spread 0.31%/mo; FF3 alpha 0.52%/mo, t=4.49): https://www.etf.com/sections/index-investor-corner/swedroe-exploring-profitability-factor
• Frazzini & Pedersen (2014), "Betting Against Beta": https://papers.ssrn.com/sol3/papers.cfm?abstract_id=1723048 (low-beta decile Sharpe ≈0.70 vs high-beta ≈0.28, 1926–2012, as summarized from the study's data)
• Harvey, Liu & Zhu (2016), "…and the Cross-Section of Expected Returns" (t ≥ 3.0; 316 factors): https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2513152 ; summary: https://www.cxoadvisory.com/big-ideas/taming-the-factor-zoo/
• Bailey & López de Prado — Deflated Sharpe (2014); Bailey, Borwein, López de Prado & Zhu — Probability of Backtest Overfitting / CSCV (2017): https://github.com/tradingstrategy-ai/docs/raw/6f3d52af051821774a647395501bac9f73673486/articles/bailey-borwein-lopez-de-prado-zhu-probability-of-backtest-overfitting.pdf
• McLean & Pontiff (2016) publication decay (~26% out-of-sample, ~58% post-publication) — as catalogued in: https://www.jedsonpinto.com/ai-radar.html
• Medallion figures (≈66% gross / ≈39% net, 1988–2018): Zuckerman, The Man Who Solved the Market (2019), widely reported.

All performance figures above are attributed to their study/sample and are gross of real-world frictions unless stated. None of them is a forecast for this engine; §3's haircuts apply.