Deep research: from IC 0.017 / net Sharpe 0 to net-of-cost performance
2026-07-22. 104-agent research run: 22 sources, 102 claims extracted, 25 verified by 3-vote adversarial panels → 20 confirmed, 5 refuted. Confidence labels per finding.
Diagnosis (high confidence — top-journal sources, unanimous votes)
- Quark's outcome is the textbook outcome, not a Quark bug. ML equity alpha
concentrates in hard-to-arbitrage stocks (microcaps, distressed, high-vol states).
Liquid S&P 500 large caps are structurally the low-gross-alpha segment:
excluding microcaps kills most of it (Avramov-Cheng-Metzker, Mgmt Sci 2023);
96% of trading-friction anomalies fail |t|=1.96 with NYSE breakpoints +
value weighting (Hou-Xue-Zhang, RFS 2020).
- The cost-agnostic two-step workflow (predict gross returns → form portfolios)
generically fails net of costs — training loads on fleeting reversal-type
features, turnover eats the spread (Jensen-Kelly-Malamud-Pedersen, RFS).
Canonical example: momentum 0.64%/mo gross, 63bps/mo cost, 0.01%/mo net.
Quark's 68x/yr turnover vs 16bps/wk spread is a direct instance.
The remedy, ranked by verified effect size
The surprise: cost-awareness in CONSTRUCTION, not new data, is the documented lever.
- Garleanu-Pedersen partial rebalancing (JF 2013). Trade partially toward an
"aim portfolio"; weight persistent (slow-decay) signals more. ~+20% net Sharpe
vs best static policy. Closed form under quadratic costs; under proportional
costs → no-trade bands. [3-0 x3, verified against author PDF]
- TCA factors (Baldi-Lanfranchi 2024 WP). Factor-specific optimal rebalance
speed → net max squared Sharpe up to 2.5x; momentum costs 63→30bps/mo,
+0.19 net Sharpe, same signal. [3-0; unrefereed magnitudes]
- Cost term inside the training objective. TC-penalized conditional
autoencoder: OOS R² 0.03%→0.27% ex-microcaps (Jo-Kim-Shin 2025 WP).
Full Portfolio-ML (learn weights directly under cost-aware objective):
~+20% net Sharpe over even a cost-aware two-stage (JKMP, RFS). [3-0 x6]
- WARNING — ex-post overlays mostly don't work on ML signals.
Azevedo-Hoegner-Velikov: holding-period extension to 2 months was the ONLY
overlay improving net performance across all 9 ML strategies tested, and only
+5bps. Buy/hold bands, quintiles, size filters: turnover falls, gross falls
with it. Implication: Quark's no-trade-band study (0.13→0.18) is directionally
consistent but bands alone won't reach the target; the cost logic must enter
the objective/rebalancing rule. [3-0, single strong source → medium]
Realistic ceiling (medium confidence, single source AHV)
Post-2005 liquid US equities, monthly ML combinations of 320 published anomalies:
OOS monthly R² 0.05–0.76%; gross VW decile Sharpe 0.32–1.11; costs 19–26bps/mo
at 120–140% monthly turnover (HF effective spreads, value-weighted); only best
sequence models (LSTM) survive net: 1.42%/mo (t=3.99).
Net Sharpe ~1 is roughly the documented frontier. IC 0.04+/net Sharpe >1 on
S&P-500-only is AT or BEYOND anything published. Live scholarly dispute:
Avramov et al. say costs kill ML strategies; AHV say good spread estimates
rescue the best ones. Truth depends on execution quality.
Refuted in verification (do NOT rely on these)
- "Value/quality features become top-importance under cost-aware training →
fundamentals are the highest-value data add" — REFUTED 0-3.
- AQR "real-world costs are many times smaller than academic estimates" — 1-2.
- AQR "size/value/momentum survive costs at very large scale" — 0-3.
- "Short-term reversal can't be rescued by optimization" — 0-3.
- "Cost-adjusted model Sharpe 1.071 vs 0.758 ex-microcaps" — 0-3.
Open questions (research needed — we generate our own evidence)
- Incremental IC per data family (fundamentals, IBES revisions, short interest,
NLP) — NO claim survived verification. Unverified leads from fetch stage worth
testing in-house once WRDS clears: PEAD (rank-SUE coeff 0.070, t=7.16,
1983-2001), post-announcement analyst revisions (3.43% 90-day CAR spread vs
2.18% for SUE, 1995-2019), short-interest surprise (SUSIR, JFM 2023),
earnings-call NLP (IC ~0.017 standalone, S&P Global 2018).
- Can GP-style rebalancing alone lift a 16bps/wk gross spread to net Sharpe ~1,
or is new data strictly necessary? (Both levers likely needed.)
- Does Quark's t=3.35 deflate under proper trial counting? (Trial registry
exists; apply DSR to the xsec family.)
Prioritized action plan for Quark
| # |
Intervention |
Documented gain |
Effort |
Data needed |
| 1 |
GP-style partial rebalancing toward aim portfolio (tune speed τ OOS); slow-signal tilt |
~halve costs, +~0.2 net Sharpe |
days |
none (have it) |
| 2 |
Retarget to longer horizon / slower components (2-mo class; Quark's own 3M IC 0.0291 > 1W 0.0172) |
+5bps (only overlay that survives) + slower decay |
days |
none |
| 3 |
Cost term in training loss (TC-penalized objective), then Portfolio-ML-style direct weight learning |
R² 0.03→0.27% ex-microcaps; +20% net Sharpe |
weeks |
none |
| 4 |
New data families via WRDS: IBES revisions, PEAD, short interest, fundamentals |
unquantified (open question) — test in-house |
weeks |
WRDS |
| 5 |
Recalibrate target: net Sharpe 0.5–0.8 on S&P-only would already be publication-grade |
honesty |
— |
— |
Full verification transcript: task output wkpy1he1v; journal at subagents/workflows/wf_a00ff121-685.