← 06 RESEARCH / KNOWLEDGE_BASE.md

Quark Knowledge Base

The consolidated brain: every verified finding, dead end, formula, and rule from the 2026-07 research campaign. Primary sources in reports/deep_research_*.md; empirical trial log in RESEARCH_NOTES.md. Confidence flags: [V] adversarially verified (3-vote), [E] our own empirical result, [U] unverified lead.

0. The one-paragraph thesis

Edges are inventory, not assets — they decay after publication (58% in US large caps, worst where arbitrage capital concentrates). An individual's only durable advantage is capacity: edges too small ($50k–$2M) for institutions to touch. The job is a pipeline: find candidate edges, validate them cheaply and honestly (PIT + walk-forward + shuffled control + efficient-market null), size survivors as if the estimate is wrong (fractional Kelly), and re-hunt as they die. The machine that runs this loop is worth more than any single edge.

1. VERIFIED EDGES (with effect sizes)

Cross-sectional / factor

[E] Cross-asset relative value is the pond. Ranking 77 instruments across 7 asset classes against each other (not time-series) → walk-forward net Sharpe 0.85, shuffled-control clean, 12/15 positive years, +0.08 corr to equity. The edge lives outside the most-arbitraged universe.
[V] Gross profitability (Novy-Marx) survives in large caps: 26bp/mo (t=1.88) in the largest quintile, incremental to value; the two are −0.57 correlated so a 50/50 mix ~doubles Sharpe (in-sample 0.74; halve it for post-decay estimate).
[E] Garleanu-Pedersen partial rebalancing (trade a fraction τ toward target, tilt to persistent signals) delivered the literature's promised +~0.2 net Sharpe on the PIT panel. Cost side is solved; the binding constraint is gross signal.
[E] Composite IC adds in quadrature: √(IC₁²+IC₂²). Adding one uncorrelated IC-0.02 signal beats improving the existing signal 20% — new families dominate. Broad-panel composite (price+insider+overnight) reached IC t=3.9.

Prediction markets (the chosen niche)

[V] Favorite-longshot bias is universal, worsens with horizon: calibration slope 0.99 (≤1h to resolution) → 1.32 (>1 month). Strongest in politics. → buy the favorite side at medium horizon; the crowd underprices near-certainties.
[V] Sub-10¢ longshots lose buyers >60%; contracts >70¢ earn positive post-fee returns. Never hold lottery tickets; the favorite side is where net-positive lives.
[V] ~$40M arbitrage extracted (Apr'24–Apr'25) from mispriced mutually-exclusive sets (buy-all <$1 / sell-all >$1). Median depth ~15 shares — retail-only by construction.
[E] The telegraph: longshot surprises that resolved YES showed a 3–4 day price staircase before the jump (e.g. 0.04→0.14→0.24→0.82). A rising longshot is informed money arriving. Read the slope, not the level. (News-shock surprises give no warning — uncatchable from odds.)
[V] Always be the maker: makers structurally out-earn takers; Polymarket pays daily quadratic LP rewards S=((v−s)/v)²·b for tight two-sided quotes.
[E] My edge = semantic verification, not information. Naive date-ladder arb scans are ~90% false positives (survival-ladders, snapshot vs cumulative, different-event). The value is reading resolution rules to filter them — a language strength, immune to my Jan-2026 knowledge cutoff.

Timing / regime

[E] Turn-of-month lives (+0.5 Sharpe, 58yrs post-publication) — loser is the payroll calendar, a structural counterparty no fund can arbitrage. Usable as a free execution-timing overlay.
[E] Storms are learnable — but NOT tradable: OOS AUC 0.76, 2.6× lift, 11/11 years (real screen). RETURN GATE FAILED (2026-07-24): top-5% states net +5.5% vs baseline +5.9%, median negative, 10.6% lose >50% — a lottery bag with no premium. Status: scanner in the drawer; classification ≠ strategy, proven on itself.

2. FALSIFIED — do not re-run without new construction

[E] S&P cross-sectional flagship: IC real (t=2.2–2.4) but walk-forward config selection → net Sharpe ~0.05. Economic edge indistinguishable from zero.
[E] Monthly S&P signal was 100% survivorship (PIT IC = 0.0000).
[E] XBRL-timed PEAD/SUE: features subtract (both horizons). 10-Q filed dates lag the 8-K announcement 2–4 weeks; drift already spent. Needs announcement dates.
[E] 8-K micro-cap earnings drift: extinct at all sizes (small-cap −0.05%, t=−0.40, n=11.8k). Machines read every filing now — the barrier commoditized.
[E] Crypto cross-sectional momentum: IC real (t=1.99 clean) but extremes mean-revert → net negative every τ.
[E] WSB attention: clean null (t=0.15). Edge died Jan-2021 (member explosion); data is adversarial (users post fake tickers to poison scrapers).
[E] Broad-universe price-only economics: IC t=3.11 but gross spread thinner than S&P; extremes reverse. Signal ≠ money.
[E] Realized weather → nat gas: fully priced (t=0.44 vs control). The edge is upstream in forecast errors, not realizations.
[V] Most published anomalies: 6 of 8 tested dead or inverted on our panel; the crowded ones (reversal, low-vol) became anti-edges that pay the other side.
[E] FLB harvesting on Polymarket (2026): FALSIFIED as-implemented. 55-trial campaign (Jan-Feb steer, Mar-Jul test): frozen config −$2,111 at MC pct 2.4% (significantly below the efficient-market null); best of 50 variants = breakeven. Stretch strength monotonically WORSE; politics the poison category. The 2021-25 documented bias is compressed/dead in 2026 tail markets — the institutional-entry edge-death clock, confirmed on our own data. Cluster caps robustly saved money in every variant (permanent doctrine). CLV called the verdict from month one. STILL-LIVE untested families: telegraph/slope momentum, pure arb, maker rewards, whale-following — all queued for live-forward paper (Aug+), the only clean data.

3. THE MATH (verified formula sheet)

[V] Fundamental law, corrected: IR = IC/√(1/(φN)+σ²_IC), capped at IC/σ_IC as N→∞. Naive IR=IC√breadth overstates wildly. Our IC 0.0125, t=2.35 ⇒ σ_IC≈0.146 ⇒ gross IR ceiling ≈0.62, effective breadth ~45–50 independent bets/wk (not 470). IC volatility bounds you as hard as mean IC — stabilizing IC (ensembling, horizon-blend, family diversification) is worth as much as raising it.
[V] Transfer coefficient: IR = TC·IC·√BR; constrained books capture only 30–80% of theoretical IR. Decile books discard information; rank-linear + inverse-vol construction recovered +34% Sharpe where gross edge existed.
[V] Kelly (continuous): f = μ/σ² single-asset; F = Σ⁻¹(μ−r) multi-asset. Fractional Kelly is mandatory: overestimate edge by 2× → full-Kelly gives negative growth. Benter ran ½–⅓ Kelly. Empirically (our 50-variant game) median wealth peaked at 1.25× leverage and fell at 1.5× — the inverted-U is real.
[V] Publication decay: 58% in US large caps, worst in liquid/large/low-idio-vol names. Haircut every published effect ~half before believing it.
CLV (to build): beating the closing price predicts long-run profit better than short-run ROI; converges in hundreds of bets, not thousands.

Betting-backtest methodology [V] (reports/deep_research_backtest_method)

[V] Sharp-reference-price / CLV is the gold standard: EV vs a sharp book's margin-removed price predicts realized returns ~1:1 (Buchdahl, 37k matches); soft prices carry no info. Our CLV-analogue = entry price vs later pre-resolution price.
[V] Significance gate = Monte-Carlo efficient-market null, NOT a t-test: simulate outcomes drawn from market-implied probabilities; a record is real only if it beats that distribution (Kaunitz 10.82σ template). Per-period testing gives 78% family-wise false positives — test the full period, correct for variants.
[V] Documented FLB frequently does NOT survive net of spread+fees: soccer/tennis every odds decile pays <1 per unit at closing odds. Calibration bias ≠ profit. The report card requires BOTH Monte-Carlo-null-significant net return AND positive CLV.
[V] Entry realism, conditioned on price level (Polymarket order-book): median spread ~400bps at 0.4–0.6, 1,300–1,800bps below 0.10, ~53bps half-spread above 0.90. Longshots are expensive to trade — the FLB "avoid <10¢" rule is also a spread rule. Model executable depth (~15 shares many episodes).
[V] Convergence contamination: exclude near-resolution phase (prices mechanically converge; MMs withdraw quotes). Entry-timing cutoff pre-specified, ≥ a few days out.
[V] Cluster-robust inference: cluster SEs at event level (mutually-exclusive contracts negatively correlated) AND contract level (repeated daily obs); never double-count Yes/No sides. Enforce per-event exposure caps.
[V] Calibration regression: use normalized probabilities when an overround exists (raw inverse-odds regression is biased against detecting FLB).
[V] Low power is severe: even 14 seasons detects a real |β|=0.05 bias <50% of the time. Expect to need many bets; a null is not proof of no edge.
Model-vs-market threshold (Kaunitz rule): bet only when model prob ≥ ~1.4× market implied — monotonically improves returns, though often not past breakeven.

4. METHODS (the gates — non-negotiable)

PIT correctness: features known on the date they appear. Value first visible the trading day after filed date. (Cycle-3 lesson: 10-Q vs 8-K timing.)
Purged walk-forward for model training (purge ≥ horizon).
Walk-forward CONFIG selection: pick τ/blend/threshold on trailing data only, measure forward. This is what took 0.28 → 0.05. The harshest, truest gate.
Shuffled-label control: real signal → shuffled ≈ 0. Always run it.
PIT/survivorship measurement: quote survivorship as a measured margin, not a disclaimer. Current-members panels are upper bounds.
Trial counting / DSR: every evaluated variant registers; quote plateaus not maxima. Selection tax measured at 98% (max vs median of 50 variants).
Efficient-market Monte-Carlo null (betting): simulate histories where prices were exactly right; a record is real only if it beats that distribution.
Data-artifact paranoia: every artifact flatters (LUNA ticker-reuse gave a fake +17M% day hiding a negative strategy; convergence-contaminated trades gave a fake step-function). Cleaning always reveals something smaller and truer.

5. MASTERS' DOCTRINE (Thorp / Benter — verified)

Thorp loop: idea → quantitative development → cheap real-world verification. (We reinvented this as the cycle loop.)
Benter: 10 man-years to build the edge (5 model, 5 operation) — the effort was the moat. Published his model and still won: some moats survive disclosure.
Blend your estimate with the market's (Benter's key technical insight): a raw model is biased vs public odds; combine them. Directly applies to bias-correcting market prices rather than forecasting independently.
Capacity-first pond selection: Benter targeted the largest pools because edge capacity = 0.25–0.5% of pool turnover. Pick the pond by its capacity.
The ladder: gambling→markets→scale. Niches are rocket fuel early, become rounding errors as wealth outgrows their capacity (our edge-factory sim reproduced this: niches 24% of profits early, 6% late). Buffett's confession: small sums earn the highest returns.

6. PREDICTION-MARKET OPERATIONS

Venues: Polymarket (CLOB on Polygon, USDC, UMA oracle — 2h challenge window, $750 bonds, 50/50 "Unknown" settlement risk; US-person restricted). Kalshi (CFTC-regulated, US-legal, taker fee 0.07·p·(1−p)).
Free data: Gamma (markets, no key), CLOB (order books), data-api (per-wallet positions/trades/leaderboard — the RIGHT tool for whale-following; the /trades endpoint is recency-first, causing convergence contamination).
Standable-on code: Polymarket/agents (AI/RAG forecaster, MIT, 3.7k★), warproxxx/poly-maker & official poly-market-maker (LP-reward makers). Scams flagged: "guaranteed YES+NO<$1" bots; Trum3it/polymarket-arbitrage-bot = malware. Real code disclaims profit.
Institutions: SIG desk 2023 / Kalshi MM Apr-2024 (NOT 2021 — corrected); DRW/Wintermute/IMC hiring 2026 = edge-death clock ticking. Liquid-sports single-market arb DEAD (7 episodes/mo, 3.6s lifespan). Capacity floor verified: combinatorial arb alive at 101bps but 77% of episodes can't fill $100.
Who loses: partisan/emotional bettors, hedgers, attention traders. Markets ≤ polls on election eve since 1936 — not an oracle.

7. OPERATING RULES (the constitution)

Sizing: portfolio vol target 10–12% (~⅕ Kelly of shrunk edge); per-niche positions ≤ ⅓ Kelly of estimated edge; theme-cluster caps (the Iran lesson: 3 markets = 1 bet); STOP-file kill switch.
Niche selection (rank order): structural loser > nameable barrier > event frequency > machine-testable > capacity fit > differential info. Anti-signals: needs latency, needs paid data, popular on FinTwit, loser is smart money.
Never: trade what hasn't passed its gates; size past policy; touch a venue you're not eligible for; quote a max instead of a plateau; trust an artifact.
Always: log every trial; measure survivorship; run the shuffled control; haircut published effects ~50%; treat a worse number on the second look as true.

8. OPEN / QUEUED

[CLOSED 07-24] Storm gate: failed — scanner only. Anti-FLB harvest: dead (−21.7%/bet net; bias real at slope 0.87 but toll bigger). 2026 center is OVERCONFIDENT (reversed from 2021-25 lit). ALL tape roads → the maker.
Prediction-market: TRUE tail calibration (gamma pagination blocks <$100k vol — needs per-day fetch); insider redo on OLDEST-first trades; wallet-follow (Desk shadow-book accruing, maker-noise filtered).
THE DESK (moneymaker3000): arb-sim 30min / maker-sim + shadow 2x daily — verdicts: arb+maker ~2wks, shadow ~6-8wks.
Forecast-vintage collector (NOAA) — the weather-error moat.
Sept: WRDS/IBES announcement-dated earnings (the family falsified only for timing).
The Jan–Jul 2026 prediction-market betting backtest (building on methodology research).

Update this file when a cycle resolves. It is the model's memory and the operator's manual. Honesty of this document = value of everything downstream.