Loop Picks — Validation Record

An honest, durable account of how the stock model was stress-tested, what survived, and what didn't. Pages: /analysis (5yr rigorous), /deep (14yr + 20yr deep history), /review (independent-LLM red-team, full transcripts).

What the strategy is

A trust-weighted ensemble where each source (YouTubers, newsletters, per-firm analysts, per-member congress, insiders, FinTwit, hype, gov contracts, lobbying, patents) is graded on its own signed alpha. The signal is convergence: a name where ≥3 distinct trusted sources spanning ≥2 independent types (echo-collapsed) turn bullish within 21 days. The trade: buy on convergence, hold winners (no profit cap), −30% catastrophic stop, 6-month max hold, ~10% slices, ≤20 positions filled by conviction-priority.

The backtests (vs SPY, QQQ, Russell 1000)

• 5yr media era (2021–26), walk-forward PIT trust + 0.3% costs + next-day fills: +250%. Per-trade vs SPY +7.6%, P(>0)=95.6%.
• 14yr cross-type, durable signals only — no media (2012–26: analyst firms + congress + price-momentum): +1,738%. Beats SPY/Russell 100% per-trade; beats QQQ +2.7%/trade, P(>0)=95.4%.
• 20yr momentum-only (2007–26, incl 2008): +3,914% (−57% drawdown = the 2008/09 momentum crash). Beats QQQ +3.0%/trade, P(>0)=95.2%.

What we found, in order

1. A real fragility (found and fixed)

The signal emitted ~1,935 candidates for a 12-slot book → ~49% dropped by arrival timing. The same strategy swung +45% to +572% across input orderings. Fix: fill slots by conviction (deterministic). The number stops being arbitrary. Locked config: ≥3 types, conviction-priority, 20-slot book.

2. The placebos — and the correction that mattered

• Within-universe placebo (random from the curated universe): real strategy at the ~65–86th percentile (p≈0.14–0.33). Within-universe timing adds little.
• Total-market placebo (random from the whole market): real at the 100th percentile — but SPY and QQQ are also at ~99th, because the average individual stock badly lags the cap-weighted indices. So "beats random-market" is a weak bar everyone clears.
• Conclusion: the apparent edge is universe construction (which stocks the signals converge on), not timing — and the right benchmarks are the indices, not random-market.

3. The decisive test — Fama-French 5 + Momentum regression

Regressing the strategy's monthly returns on Mkt/SMB/HML/RMW/CMA/MOM: alpha = +2.5%/yr, t = 0.87. Not significant. Loadings: Mkt 1.02 (t=17), MOM +0.27 (t=3.8) — a market portfolio with a momentum tilt; the other factor loadings are small/insignificant. The alpha is positive but not statistically significant — unproven, not disproven. We can't claim a real edge; we also can't claim its absence.

Correction (2026-06-23): an earlier version of this page reported −0.8%/yr, t=−0.14, "no alpha beyond factors." That came from an OLS bug (a missing matrix multiply, (XʹX)⁻¹y instead of (XʹX)⁻¹(Xʹy)) that returned near-zero garbage betas and systematically biased every alpha test toward zero. Fixed; the regression now correctly shows market beta ≈ 1. The bootstrap-based results elsewhere on this page never used OLS and are unaffected.

• Deflated Sharpe (Bailey & López de Prado, ~30-trial haircut): annualized Sharpe 1.01, PSR-vs-0 = 100%, deflated = 92.5% (marginal, just under 95%).

4. Signal-type independence (measured)

Mean off-diagonal co-occurrence lift 1.90× ("mildly correlated"). Media types are redundant (Talking Heads ↔ Newsletter ~12–15×), but the hard-data types are genuinely independent (Insider ~0 lift with Analyst/Congress/Patents).

5. Independent-LLM red-team (full transcripts at /review)

Six frontier models, told to be hostile, scored "does it have alpha?" at mean 2.8/10. Recurring valid criticisms: survivorship (universe and sources), universe-does-the-work, data-snooping, factor contamination.

• Calibration control: the same panel scored Jegadeesh-Titman momentum (a proven 1993 factor, t>4) at 2.8/10 — identical. So the absolute score is a floor for any "beat the market" claim, which the models explicitly admitted ("guilty until proven innocent").
• Re-poll (honest factor-strategy framing + bias confronted): 6.7/10.
• Discrimination control (same called-out framing): momentum 6.8, deliberate garbage overfit 0.5. So the panel does discriminate — the 6.7 is a real "legitimate factor strategy" signal, not sycophancy. But it is not "special" — momentum ties it.

The honest verdict

Claim	Verdict
Beats SPY / Russell decisively, QQQ at ~95% per-trade (5–20yr, net of costs + walk-forward)	✅ Yes
Has alpha beyond known factors (Fama-French)	— Unproven (+2.5%/yr, t=0.87, not significant)
Is a sound, investable factor-harvesting strategy (panel 6.7 = momentum, ≫ garbage)	✅ Yes
Is differentiated from cheap factor ETFs on the backtest evidence	— Not shown

The convergence logic is a transparent rebalancing rule for getting factor exposure, not a secret edge. The only open question is differentiation, and no backtest can answer it. The live forward track (running weekly, survivorship-free by construction) is the sole remaining arbiter.

Survivorship-free corpus (Sharadar) — what it found

We bought the full survivorship-free corpus (21,862 stocks incl. 15,573 delisted, 1998–2026) and built a reusable testing framework. Most signals failed rigorous testing (insider cluster-buys show negative selection; convergence is factor-explained), but one survived: institutional 13F-accumulation — +6.16%/6mo vs a size+momentum+value+quality characteristic-matched basket, P=100%, stable every era, and it generalizes across size. The skeptical LLM panel moved from "factor mirage" to candidate edge, 6.8/10. Honest bound: this rests on one rigorous method (matched-attribution); the regression cross-check was abandoned (invalid for clustered events), not faked. A genuine candidate, not a confirmed edge. Full detail and transcripts: the corpus audit & model validity page →