WEEKEND RESEARCH
PremiumResearchJournal

Weekend Research Drop

Updated Oct 23, 2026 Β· 11:00 PM Β· organized by verdict β€” archive collapsed at bottom

πŸ† The Verdict β€” what survives rigorous testing

Start here. Every claim below is re-tested with multiple-comparison control, bootstrap CIs and regression β€” this is what the whole research program nets out to.

What Actually Survives β€” FDR Correction Across the Whole Arc

We have tested 19 major claims across two months of research. Test enough things and some look good by luck. This recomputes every claim's p-value from source events and applies Benjamini-Hochberg false-discovery control (q=10%). Claims that fail here should not be traded, no matter how good the headline looked.

Claimnmeantpsurvives FDR
Buffett 13F new stake (+5d)67+3.01%4.990.0βœ…
SP500 earnings reaction (+1d)23,128+0.26%6.460.0βœ…
Top-100 13F stakes (+1d)96,473+0.08%6.40.0βœ…
FDA approval (+20d)2,463-0.98%-4.272e-05βœ…
Insider cluster BIG-$ (top-half value) (+20d)1,458+1.56%3.370.00077βœ…
Gov-contract award (+20d)275-2.66%-3.380.00083βœ…
Insider cluster RISING (+20d)827+2.14%3.250.0012βœ…
Super-investor 13F stakes (+1d)7,876+0.10%2.970.00297βœ…
Insider cluster w/ OFFICER (+20d)3,221+0.78%2.580.01001βœ…
Insider cluster LATE filers 6d+ (+20d)184-3.21%-2.480.01393βœ…
52w-high breakout (+20d)8,005+0.18%2.270.023βœ…
Insider RISING Γ— lead-sector (+20d)350+2.72%2.240.02579βœ…
Insider cluster β€” all (+20d)3,594+0.59%2.050.04042βœ…
PEAD big beat (+60d)49+6.69%1.710.09394❌
Insider cluster FALLING/'dip' (+20d)1,588+0.75%1.650.09858❌
Insider cluster ON-TIME filers ≀2d (+20d)2,316+0.37%1.080.28004❌
Insider cluster directors/10%-only (+20d)373-1.03%-1.080.28179❌
Insider cluster LIVE feed (+20d)643+0.18%0.480.63301❌
Activist 13D (+20d)79-0.17%-0.130.89605❌

13 of 19 claims survive false-discovery correction.

Read this

  • βœ… rows are the firm's defensible research conclusions β€” everything else is either noise or too weak to distinguish from it after multiple testing.
  • This is deliberately the harshest single table on the site. If a strategy's premise fails here, the strategy needs a different justification (or retirement).

Flagship Strategy P&L Simulation β€” what to actually expect

If the live bot trades ONLY the validated RISINGΓ—leading-sector clusters (~33 trades/yr historically, 10bp costs), the simulated annual ALPHA contribution per unit of strategy capital (20,000 bootstrap years):

Per-trade allocp5 (bad yr)medianp95 (good yr)P(positive yr)
5.0%-6.09%+4.15%+15.41%74%
7.2%-8.66%+5.99%+22.40%75%
10.0%-12.35%+8.23%+30.82%74%

Read this

  • These are ALPHA numbers (excess over SPY), additive to market return.
  • Median is what a typical year looks like; p5 is the bad-luck year to size for.
  • ~51% win rate + positive-skew means dry spells are normal β€” the p5 row is the emotional preparation, the median row is the plan.

Deep Statistical & Probability Analysis

Rigorous inference on 27,511 signal events (+20d outcome), with a new volatility+liquidity factor layer. Bootstrap=10,000 resamples, permutation=5,000.

1. Per-signal edge β€” bootstrap 95% CI (is it real?)

Signalnmean +20d95% CIwin%binom psignificant
activist-13d79-0.17%[-2.40%, +2.63%]43%0.26043β€”
buffett-13f67+3.01%[+1.90%, +4.20%]75%7e-05βœ…
insider-cluster3,594+0.59%[+0.03%, +1.15%]48%0.01293βœ…
insider-cluster-live643+0.18%[-0.55%, +0.90%]53%0.20693β€”
sp500-earnings23,128+0.26%[+0.18%, +0.34%]52%0.0βœ…

2. Logistic regression β€” what drives P(win) controlling for everything

Significant factors only (vs reference level), ranked by p-value. Odds ratio >1 = raises win prob.

Factorodds ratiozp
signal_type=insider-cluster0.474-4.980.0
liq_bucket=MID1.1594.630.0
prior_trend=RISING1.943.970.0001
prior_trend=FALLING1.7613.610.0003
trend=BEAR1.1463.620.0003
prior_trend=FLAT1.7683.460.0005
signal_type=buffett-13f2.6913.350.0008
vix_bucket=STRESS0.902-2.640.0082
liq_bucket=SMALL1.2092.450.0144
signal_type=activist-13d0.559-2.080.0374
sector=Communication Services1.2192.070.0386

Reference levels: signal_type=sp500-earnings, sector=Unknown, prior_trend=, trend=BULL, vix_bucket=NORMAL, vol_bucket=LOW, liq_bucket=LARGE

3. Conditional win-probability by factor (nβ‰₯30, binomial vs 50%)

sector β€” significant levels: Communication Services 56% (p=0.0014), Industrials 52% (p=0.01073), Information Technology 54% (p=1e-05)

prior_trend β€” significant levels: 51% (p=1e-05)

trend β€” significant levels: BEAR 53% (p=0.0), BULL 51% (p=0.0479)

vix_bucket β€” significant levels: NORMAL 52% (p=0.00026)

vol_bucket β€” significant levels: LOW 51% (p=0.00552), MED 51% (p=0.00588)

liq_bucket β€” significant levels: MID 54% (p=0.0)

4. Interaction β€” RISING Γ— leading-sector insider clusters

  • Combo filter: +2.47% [95% CI -4.18%, +11.79%], n=7 β€” not significant
  • Insider baseline: +0.53% [+0.04%, +1.03%], n=4237

Read this

  • A signal whose bootstrap CI excludes zero is a real edge; if the CI straddles 0, the apparent mean could be noise regardless of how big it looks.
  • The logistic regression is the key: a factor that stays significant there is driving the outcome even after controlling for sector/regime/vol/liquidity β€” that's a genuine, non-confounded effect.
  • If RISINGΓ—lead-sector survives the permutation test, the combined insider filter is the firm's most defensible equity edge.

Combined Insider Filter β€” validated with full sector coverage

Sector resolved for 96.5% of 3,594 priced historical insider clusters (was ~SP500-only before). Now the RISING-trend Γ— leading-sector filter can be tested properly. Bootstrap=20,000, permutation=10,000.

The combined edge

  • RISING Γ— leading-sector: +2.72% [95% CI +0.38%,+5.05%] n=350 win 52% βœ…SIG
  • RISING only: +2.14% [95% CI +0.85%,+3.42%] n=827 win 51% βœ…SIG
  • Leading-sector only: +1.23% [95% CI +0.16%,+2.33%] n=1,452 win 49% βœ…SIG
  • Insider baseline (all clusters): +0.59% [95% CI +0.02%,+1.17%] n=3,594 win 48% βœ…SIG
  • Permutation vs baseline: p = 0.0112 β†’ beyond chance βœ…
  • Combo walk-forward (no lookahead): +2.44% / 51% (n=333)

Prior-trend Γ— leading sector

Prior trendedge
RISING+2.72% [95% CI +0.40%,+5.16%] n=350 win 52% βœ…SIG
FLAT+1.30% [95% CI -1.17%,+3.88%] n=227 win 49% ns
FALLING+1.29% [95% CI -0.22%,+2.79%] n=742 win 48% ns

Insider-cluster edge by sector (nβ‰₯25, full coverage)

Sectoredge
Healthcare+1.82% [95% CI +0.17%,+3.50%] n=778 win 49% βœ…SIG
Consumer Cyclical+1.32% [95% CI -0.80%,+3.48%] n=249 win 49% ns
Utilities+1.27% [95% CI -0.81%,+3.79%] n=103 win 49% ns
Energy+0.94% [95% CI -1.03%,+3.01%] n=228 win 45% ns
Communication Services+0.92% [95% CI -2.43%,+4.46%] n=123 win 45% ns
Industrials+0.88% [95% CI -0.87%,+2.67%] n=359 win 49% ns
Basic Materials+0.56% [95% CI -2.45%,+3.71%] n=131 win 48% ns
Consumer Defensive+0.55% [95% CI -1.98%,+3.16%] n=113 win 47% ns
Real Estate+0.40% [95% CI -1.07%,+1.91%] n=234 win 49% ns
Financial Services-0.02% [95% CI -0.87%,+0.84%] n=822 win 48% ns
Technology-0.83% [95% CI -3.15%,+1.52%] n=270 win 47% ns

Read this

  • If RISINGΓ—leading-sector has a CI excluding 0 AND beats the baseline AND survives the permutation + walk-forward tests, it's the firm's most defensible, validated equity edge β€” the sizing/filter to put on the live insider bot.
  • If it only matches RISING-alone, sector adds nothing beyond the prior-trend signal.
  • Sectors whose CI excludes 0 are where insider clusters genuinely work.

πŸ“Œ Validated Signal Studies

The signal research behind the verdict β€” deep samples, walk-forwards, live watches.

Insider Cluster β€” Deep Historical Backtest (2016-2026)

Scraped 5,153 cluster-buy events from openinsider (>=2 distinct insiders, same ticker, within 30 days, purchases only) and priced 3,594 with yfinance abnormal returns (ticker minus SPY). This is the deep sample behind the live bot, which until now rested on ~63 clusters.

Overall edge

HorizonnWin%Mean abnMedian
+5d3,59149%+0.68%-0.09%
+20d3,59448%+0.59%-0.52%
+60d3,42542%-2.37%-3.36%

By cluster size (+20d) β€” does more conviction pay?

# insidersnWin%Mean
22,24547%+0.66%
380649%+0.76%
429950%-0.11%
5+24452%+0.29%

By year (+20d) β€” is the edge stable or fading?

YearnWin%Mean
201623250%+0.90%
201725250%+0.62%
201824055%+2.94%
201929143%+0.81%
202032253%+2.79%
202131750%-0.27%
202233149%-1.09%
202337239%-1.08%
202441338%-1.66%
202547954%+2.86%
202634549%+0.26%

NEW β€” by prior trend (buy the dip vs buy strength?)

Stock's own 21-day return BEFORE the cluster. Tests whether insider clusters work better as contrarian (buying weakness) or momentum signals.

Prior trendnWin%Mean +20d
FALLING1,58849%+0.75%
FLAT87248%+0.71%
RISING82751%+2.14%

Embargo walk-forward (honest, no lookahead)

  • +20d gated: n=3,524 Β· +0.54% Β· 48% WR vs full +20d +0.59% / 48%.

Read this

  • This is the real foundation the live insider-cluster bot needed. Compare the deep mean to the ~63-cluster sample it was launched on.
  • A monotone size effect (3 > 2, 4 > 3) justifies a conviction filter.
  • A by-year mean that holds post-2022 means the edge isn't arbed away. A fading tail is the warning sign.
  • openinsider caps 1000 rows/quarter, so the highest-volume quarters (e.g. COVID Q1-2020) are under-sampled toward their latest filings.

Flagship Filter Tuning β€” regime & hold-period

Two tuning layers on the validated RISINGΓ—leading-sector filter (n=356).

Hold-period (which exit?)

Horizonedge
+5d+1.41% [CI +0.17%,+2.70%] n=355 win 53% βœ…
+20d+2.72% [CI +0.36%,+5.10%] n=350 win 52% βœ…
+60d-1.48% [CI -5.27%,+2.33%] n=325 win 43% ns

Market regime (+20d)

Regimeedge
SPY BULL+3.55% [CI +0.97%,+6.18%] n=300 win 52% βœ…
SPY BEAR-2.25% [CI -8.02%,+3.25%] n=50 win 54% ns
VIX CALM+2.26% [CI -1.55%,+6.22%] n=133 win 50% ns
VIX NORMAL+4.47% [CI +1.14%,+7.86%] n=169 win 56% βœ…
VIX STRESS-2.15% [CI -8.78%,+4.71%] n=48 win 46% ns

Read this

  • The horizon with the best significant mean sets the live bot's hold.
  • If BEAR/STRESS cells are ns or negative, add a regime gate to the filter; if they hold, the edge is regime-robust (better).

Edge Stability β€” rolling 24-month decay curves

Rolling 24-month windows (quarterly steps) of the +20d mean with bootstrap CI. This is the 'is the edge still alive' view that by-year tables blur.

Window startbaseline mean [CI]flagship RISINGΓ—sector mean [CI]
2016-02-01+0.75% [-0.54%,+2.09%] n=484+9.14% [+1.14%,+17.89%] n=32
2016-05-02+0.71% [-0.46%,+1.98%] n=487+7.74% [-0.55%,+16.91%] n=33
2016-08-01+1.30% [+0.05%,+2.63%] n=488+10.69% [+3.37%,+18.93%] n=38
2016-10-31+1.05% [-0.24%,+2.40%] n=485+7.04% [+0.69%,+13.73%] n=39
2017-01-30+1.75% [+0.48%,+3.12%] n=492+6.62% [-0.64%,+14.42%] n=34
2017-05-01+1.00% [-0.29%,+2.32%] n=497+4.62% [-1.54%,+11.39%] n=38
2017-07-31+1.01% [-0.34%,+2.44%] n=491+4.86% [-1.58%,+11.87%] n=39
2017-10-30+1.34% [-0.03%,+2.76%] n=508+4.38% [-1.32%,+10.55%] n=41
2018-01-29+1.77% [+0.43%,+3.15%] n=531+3.08% [-2.59%,+8.73%] n=43
2018-04-30+1.73% [+0.21%,+3.28%] n=559+3.16% [-2.79%,+9.26%] n=36
2018-07-30+2.11% [+0.64%,+3.60%] n=588+4.68% [-1.32%,+11.10%] n=41
2018-10-29+2.09% [+0.71%,+3.58%] n=606+5.95% [+0.18%,+12.32%] n=43
2019-01-28+1.85% [+0.48%,+3.30%] n=613+6.58% [+2.06%,+11.49%] n=57
2019-04-29+2.09% [+0.71%,+3.56%] n=628+7.31% [+1.99%,+13.00%] n=56
2019-07-29+2.13% [+0.75%,+3.55%] n=639+7.38% [+2.43%,+12.54%] n=60
2019-10-28+2.24% [+0.87%,+3.59%] n=643+6.14% [+1.23%,+11.31%] n=61
2020-01-27+1.27% [-0.12%,+2.68%] n=639+5.60% [+0.37%,+11.55%] n=56
2020-04-27+0.40% [-0.81%,+1.65%] n=636+3.87% [-1.33%,+9.47%] n=62
2020-07-27-0.46% [-1.72%,+0.82%] n=630+0.81% [-4.71%,+6.46%] n=57
2020-10-26-0.85% [-2.11%,+0.43%] n=639-0.18% [-5.39%,+5.11%] n=64
2021-01-25-0.69% [-1.93%,+0.62%] n=648-2.85% [-8.33%,+2.71%] n=62
2021-04-26-0.97% [-2.19%,+0.30%] n=679-3.21% [-8.19%,+1.72%] n=60
2021-07-26-1.27% [-2.47%,-0.04%] n=690-2.00% [-7.07%,+3.10%] n=72
2021-10-25-1.97% [-3.17%,-0.74%] n=690-1.84% [-7.24%,+3.65%] n=67
2022-01-24-1.08% [-2.33%,+0.14%] n=703+0.83% [-4.30%,+6.11%] n=84
2022-04-25-0.84% [-2.08%,+0.41%] n=712+0.96% [-4.22%,+6.15%] n=86
2022-07-25-1.38% [-2.59%,-0.20%] n=741-0.07% [-4.77%,+4.84%] n=91
2022-10-24-1.08% [-2.29%,+0.08%] n=764-0.40% [-5.22%,+4.42%] n=91
2023-01-23-1.38% [-2.58%,-0.18%] n=785+0.41% [-4.45%,+5.35%] n=89
2023-04-24-1.16% [-2.38%,+0.03%] n=801+1.00% [-3.68%,+5.76%] n=94
2023-07-24-0.60% [-1.83%,+0.60%] n=824+2.38% [-2.42%,+7.37%] n=90
2023-10-23+0.10% [-1.13%,+1.30%] n=857+2.35% [-2.07%,+6.87%] n=101
2024-01-22+0.77% [-0.46%,+1.99%] n=892+1.48% [-2.77%,+5.82%] n=92
2024-04-22+1.08% [-0.12%,+2.29%] n=911+2.32% [-2.35%,+7.04%] n=100
2024-07-22+2.11% [+0.90%,+3.32%] n=926+3.57% [-0.89%,+8.09%] n=97

Baseline: early +0.92% β†’ late +1.32% β€” HOLDING/STRENGTHENING

Flagship: early +9.19% β†’ late +2.46% β€” DECAYING

Read this

  • A flagship verdict of HOLDING means the filter isn't a backtest relic β€” it has worked in recent windows too.
  • DECAYING β†’ tighten expectations and watch the live forward sample closely.

People Layer β€” who is buying matters

Joins insider NAMES, TITLES and BUY VALUES (never used before) to 3,594 priced clusters. Three questions: do officer buys beat director buys, does size-of-buy matter, and are there specific repeat-winner humans worth following?

1. Role layer (+20d)

Cluster containsedge
has CEO+0.57% / 48% win (n=2,662, p=0.0381)
has CFO+0.76% / 49% win (n=1,651, p=0.3497)
has officer any+0.78% / 48% win (n=3,221, p=0.0526)
directors 10pct only-1.03% / 45% win (n=373, p=0.049)

2. Conviction layer β€” total $ bought in cluster (+20d)

Quartileedge
Q1 smallest-0.39% / 45% win (n=882, p=0.0064)
Q2+0.08% / 46% win (n=881, p=0.0152)
Q3+1.76% / 50% win (n=881, p=0.8398)
Q4 largest+1.33% / 52% win (n=880, p=0.3283)

(quartile edges: $39,929 / $163,981 / $728,253)

3. Star-insider leaderboard (β‰₯4 priced clusters, mean>0, winβ‰₯60%)

Insidereventswinmean +20dbinom p
Cox Kevin Brian4100%+37.62%β€”
Hoyen Andrew T.4100%+28.01%β€”
Dove Ross667%+26.07%0.6875
Stetson David J.4100%+20.95%β€”
Roberts Brian Kenneth475%+20.51%β€”
Column Group III Gp, LP475%+15.43%β€”
Gomes Sanjiv4100%+15.39%β€”
Edmunds Chris475%+15.10%β€”
Duggan Robert W580%+14.95%0.375
Nijhawan Pardeep560%+14.68%1.0
Brooks Michael J560%+14.68%1.0
Morris Gabriel S560%+14.58%1.0
Rachman Ilya M560%+14.58%1.0
Dalton Travis683%+14.11%0.2188
Moran Sean F.667%+14.05%0.6875
Blackstone Group Inc4100%+13.01%β€”
Cauley Robert E888%+11.32%0.0703
Dwyer Robert J989%+10.73%0.0391
Reynolds Cameron John560%+9.99%1.0
Haas G Hunter IV1090%+9.77%0.0215
Opaleye Management Inc.771%+9.74%0.4531
Ressler Richard S475%+9.31%β€”
Shemesh Avraham475%+9.31%β€”
Kuba Shaul475%+9.31%β€”
Hogan Paul R.580%+8.67%0.375

3b. Empirical-Bayes shrunk leaderboard (honest version)

Each person's mean pulled toward the grand mean (+0.44%) with weight n/(n+k), k=2. What survives shrinkage is far more likely real.

Insidernraw meanshrunk mean
Cox Kevin Brian4+37.62%+25.61%
Dove Ross6+26.07%+19.89%
Hoyen Andrew T.4+28.01%+19.11%
Paulson & Co. Inc.4+21.93%+14.99%
Stetson David J.4+20.95%+14.33%
Roberts Brian Kenneth4+20.51%+14.03%
Duggan Robert W5+14.95%+10.94%
Dalton Travis6+14.11%+10.81%
Moran Sean F.6+14.05%+10.77%
Nijhawan Pardeep5+14.68%+10.75%
Brooks Michael J5+14.68%+10.75%
Morris Gabriel S5+14.58%+10.67%
Rachman Ilya M5+14.58%+10.67%
Column Group III Gp, LP4+15.43%+10.59%
Gomes Sanjiv4+15.39%+10.56%

Read this

  • If has-CEO/CFO beats directors-only, weight the live bot toward officer clusters.
  • Monotone value quartiles = conviction (size) is signal; flat = it isn't.
  • The leaderboard is a live watchlist of PEOPLE β€” but n is small per person; treat individual entries as candidates to forward-track, not proven edges. Multiple-testing risk is real with 1,700+ people scored (see the FDR study).

Earnings-Proximity Layer β€” do post-blackout insider buys carry more signal?

Joined 1,658 SP500 insider clusters to their issuer's historical earnings dates. Insiders are locked out pre-earnings; buys right AFTER the release are the first legal action on fresh information.

Cluster edge by days since last earnings (+20d)

Windowedge
POST 0-15d+1.33% [CI +0.08%,+2.63%] n=714 win 48% βœ…
MID 16-45d+0.30% [CI -0.81%,+1.47%] n=748 win 47% ns
LATE 46+d-0.10% [CI -2.58%,+2.48%] n=196 win 45% ns

Flagship ∩ POST-earnings: +5.17% [CI +0.55%,+9.88%] n=87 win 52% βœ…

Size split (fundamentals layer, median $0.4B β€” static snapshot caveat)

Sizeedge
Large (β‰₯ median)-0.27% [CI -1.55%,+1.06%] n=204 win 49% ns
Small (< median)+0.09% [CI -2.02%,+2.25%] n=201 win 50% ns

Read this

  • If POST 0-15d dominates, add an earnings-recency filter to the live bot and use the forward next-earnings cache to avoid entering right before a print.
  • The size split checks the edge isn't just a small-cap artifact (mcap is a today-snapshot joined to history β€” a real time series accumulates in the fundamentals layer from now on).

Execution Realism & Overnight Decomposition

Using the new Open-price layer: 3,257 clusters re-measured from the achievable NEXT-OPEN entry (what the live bot actually gets) and decomposed into overnight vs intraday components over the 20-session hold.

1. Does the edge survive real entry timing?

Cohortclose-entry (convention)open-entry (achievable)
All clusters+2.75% [CI +2.12%,+3.36%] n=3,257 βœ…+1.21% [CI +0.61%,+1.82%] n=3,257 βœ…
Flagship RISINGΓ—sector+5.26% [CI +2.79%,+7.75%] n=331 βœ…+3.18% [CI +0.80%,+5.63%] n=331 βœ…

2. Where does the return accrue?

Cohortovernight componentintraday component
All clusters+4.40% [CI +3.92%,+4.89%] n=3,257 βœ…-0.45% [CI -1.08%,+0.21%] n=3,257 ns
Flagship+6.26% [CI +4.66%,+7.87%] n=331 βœ…+0.27% [CI -2.22%,+2.81%] n=331 ns

Read this

  • If open-entry β‰ˆ close-entry, the backtest convention isn't flattering us and the live bot should capture the published edge.
  • A big overnight share means holding through closes is essential (no day-trade variant) and gaps carry the news β€” consistent with the anomaly literature.
  • Open cache is still backfilling; skipped events shrink weekly.

Filing-Lag Layer β€” does *how fast* insiders file matter?

Form 4 is due within 2 business days of the trade. The raw scrape carries both dates, but every prior study keyed only on the filing date. This tests the lag itself across 3,573 priced clusters (median cluster lag used).

Cluster edge by filing speed (+20d)

Filing speededge
FAST 0-1d+0.32% [CI -0.54%,+1.21%] n=1,350 win 47% ns
ON-TIME 2d+0.44% [CI -0.60%,+1.46%] n=966 win 49% ns
SLOW 3-5d+1.81% [CI +0.69%,+2.96%] n=1,067 win 50% βœ…
LATE 6d+-3.11% [CI -5.55%,-0.60%] n=190 win 40% βœ…
*all clusters*+0.62% [CI +0.05%,+1.19%] n=3,573 win 48% βœ…

Flagship (RISING Γ— leading sector) by filing speed

Filing speededge
FAST 0-1d+1.61% [CI -2.30%,+5.71%] n=107 win 49% ns
ON-TIME 2d+4.66% [CI +0.04%,+9.42%] n=99 win 56% βœ…
SLOW 3-5d+2.80% [CI -1.35%,+6.93%] n=117 win 56% ns
LATE 6d++0.73% [CI -8.99%,+11.89%] n=23 win 44% ns

Correlation(lag, +20d return) = -0.044 β€” near zero means the lag isn't mechanically driving measured returns.

Read this

  • A FAST-filer premium would be a free extra filter (the data is already in every live feed row).
  • A LATE-filer premium would be suspicious: long lags mean the trade is stale by the time it's public, so any 'edge' is more likely drift we couldn't have captured β€” check the correlation line above before believing it.
  • Flat across buckets = filing speed carries no information; drop the variable and stop wondering.

Abel-Era Watch β€” does the Buffett bump survive the handoff?

Warren Buffett stepped down at the end of 2025. The live buffett-13f executor trades an edge measured almost entirely on Buffett-era filings (+2.93% / 73% WR, n=84). If the post-disclosure under-reaction was a Buffett/Munger brand effect, it may fade under Greg Abel. This table tracks every post-handoff filing event as it resolves.

Post-handoff events (+5d abnormal vs SPY)

FilingTickerIssuer+1d+3d+5d
2026-02-17LLYVKLIBERTY LIVE HOLDINGS INC-0.04%+2.69%+0.29%
2026-02-17FWONBLIBERTY MEDIA CORP DEL-0.50%-0.96%-0.66%
2026-02-17NYTNEW YORK TIMES CO+1.48%+4.39%+1.79%
2026-05-15GOOGLALPHABET INC+0.11%-2.26%-4.36%
2026-05-15DALDELTA AIR LINES INC+0.08%+5.26%+7.54%

Post-handoff so far: n=5 Β· mean +0.92% Β· win rate 60% β€” vs baseline +2.93% / 73%.

Read this honestly

  • n is tiny. This is a tracking table, not a verdict β€” it takes 3-4 Abel filings (~a year) before the comparison means anything.
  • Decision rule already in config/buffett_13f.json: re-validate after 3-4 Abel-era filings; if the bump is gone, disable the executor.
  • Note the 2026-02-17 filing still reflects Buffett-era Q4-2025 decisions; 2026-05-15 (Q1-2026) is the first true Abel quarter.

Cross-Signal Aggregation & Correlations

Master table unions 29,052 signal events across 5 signal types, each stamped with two new control variables: market regime (SPY vs 200d MA Γ— VIX bucket) and GICS sector. Updated continuously.

1. Does signal stacking pay? (co-occurrence within 60d)

  • Stacked (ticker hit by β‰₯2 signal types): -1.47% mean / 43% win (n=980)
  • Single signal only: +0.39% mean / 52% win (n=26,306)
  • Verdict: stacking does NOT add edge

Most common signal pairs

Pairco-occurrences
insider-cluster-live+sp500-earnings764
insider-cluster+sp500-earnings185
insider-cluster+insider-cluster-live40
buffett-13f+sp500-earnings34
activist-13d+sp500-earnings14
buffett-13f+insider-cluster-live2
buffett-13f+insider-cluster1

2. Performance by market regime (+20d, all signals)

RegimenWin%Mean
BULL/CALM11,22751%+0.26%
BULL/NORMAL9,10151%+0.41%
BULL/STRESS1,68748%+0.32%
BEAR/CALM13352%+0.28%
BEAR/NORMAL2,12055%+0.61%
BEAR/STRESS3,00252%+0.10%

3. Insider-cluster edge by sector (+20d, nβ‰₯10)

SectornWin%Mean
Health Care6758%+2.41%
Consumer Discretionary5655%+2.32%
Communication Services2339%+2.00%
Information Technology5763%+1.92%
Materials4161%+1.32%
Industrials7949%+0.90%
Unknown328748%+0.77%
Utilities6752%+0.51%
Consumer Staples5351%-0.63%
Real Estate3847%-0.68%
Financials15349%-0.80%
Energy9046%-3.25%

3b. NEW VARIABLE β€” insider clusters by prior trend

Were insiders buying weakness or strength? (stock's own 21d return before)

Prior trendnWin%Mean +20d
FALLING1,50049%+0.86%
FLAT82348%+0.78%
RISING78351%+2.29%

Strongest filter stacked: RISING clusters by sector (+20d, nβ‰₯8)

SectornWin%Mean
Unknown76251%+2.33%

4. Signal correlation matrix (monthly mean return)

activist-13dbuffett-13finsider-clusterinsider-cluster-livesp500-earnings
activist-13d+1.00+0.14+0.16-0.01+0.04
buffett-13f+0.14+1.00+0.04-0.09+0.08
insider-cluster+0.16+0.04+1.00+0.05+0.03
insider-cluster-live-0.01-0.09+0.05+1.00+0.03
sp500-earnings+0.04+0.08+0.03+0.03+1.00

Read this

  • Stacking: if stacked > single, a 'confluence' bot (trade only names with β‰₯2 signals) is the highest-conviction play.
  • Regime: the cell with the biggest mean tells you when these signals work; near-zero or negative cells are when to stand down.
  • Sector: lean insider-cluster sizing toward its best sectors.
  • Correlation matrix: low/negative pairs are diversifying β€” combine them; high pairs are the same bet twice.

🚫 Honest Nulls & Negatives

Tested and rejected β€” published so nobody (including us) chases them again.

Activist 13D Event Study

Full initial-SC-13D history from five activists, target tickers resolved from EDGAR filing headers. 82 events with prices. A 13D is filed within 5 business days of crossing 5% with activist intent β€” a fast, high-signal disclosure vs the 45-day 13F lag.

Abnormal return around disclosure (target βˆ’ SPY)

HorizonnWin%Mean
+5d7942%-0.95%
+20d7943%-0.17%
+60d7739%-3.80%

By activist (+20d)

ManagernWin%Mean
Pershing Square (Ackman)1436%+0.35%
Starboard Value3842%+0.01%
Third Point (Loeb)1362%+1.08%
Icahn Capital1436%-2.34%

Read this

  • 13D is the firm's fastest filing signal (5-day lag). If the +5d/+20d mean is positive, a 13D-follow bot complements the slower 13F bots.
  • Per-manager N is small β€” treat manager breakdowns as directional.
  • The activist-watcher already streams these live to the dashboard; this study tells us whether to trade them, not just watch.

Government Contract Award β€” Event Study

Does a public company outperform after winning a material (β‰₯$50M) federal contract? Priced 275 ticker-tagged awards (winner βˆ’ SPY).

Overall reaction

HorizonnWin%Mean
+5d27445%-1.01%
+20d27542%-2.66%
+60d26250%-4.52%

By awarding agency (+20d, nβ‰₯5)

AgencynWin%Mean
Department of Energy956%+1.03%
Department of Health and Human3161%+0.00%
Department of the Interior650%-0.06%
National Aeronautics and Space4134%-0.10%
Department of Education1145%-0.14%
Department of Commerce540%-0.36%
General Services Administratio6242%-0.50%
Department of the Treasury850%-0.65%
Department of State757%-1.75%
Department of Homeland Securit2938%-3.18%
Department of Veterans Affairs2726%-4.33%
Department of Transportation2236%-20.53%

Read this

  • A positive +20d mean = contract wins are a tradeable catalyst; near-zero = the award is already priced in by the time it's public.
  • Sample is young (collector started 2026-06); treat as directional until it grows.

FDA Drug-Approval Event Study

Approvals 2018+ from openFDA, sponsor mapped to public ticker, priced (sponsor βˆ’ SPY). 2463 priced events. Biotech catalyst test.

Reaction around approval

HorizonnWin%Mean
+5d246346%+0.00%
+20d246341%-0.98%
+60d241249%-1.93%

Read this

  • Approvals are often pre-announced (PDUFA dates known), so the post-approval drift may be muted or negative ('sell the news').
  • Name-matching sponsorβ†’ticker is fuzzy; small/foreign pharma under-captured. Directional.

52-Week-High Breakout β€” Momentum vs Mean-Reversion

Scanned the S&P 500 universe for new 52-week-high closes (β‰₯60 trading days apart per name) and measured forward abnormal return vs SPY. 8,005 breakout events.

Forward reaction

HorizonnWin%Mean abn
+5d8,00550%+0.04%
+20d8,00550%+0.18%
+60d7,99448%+0.04%

By year (+20d)

YearnWin%Mean
201659750%+0.70%
20171,06053%+0.30%
201872453%+0.27%
201989146%-0.44%
202073045%-0.43%
20211,11245%-0.64%
202241664%+2.08%
202364146%-0.40%
202497649%+0.25%
202565847%+0.21%
202620069%+4.78%

Read this

  • Positive forward mean = breakouts trend (momentum works, buy strength) β€” consistent with the insider RISING-trend finding.
  • Negative = new highs fade (fade the rip). The +5d vs +60d shape tells you the holding window.

Sector ETF Rotation / Relative Strength

Momentum ranking of the 11 SPDR sector ETFs + a simple 'own top-3 momentum sectors, rebalance monthly' backtest vs SPY (6yr).

Current momentum ranking (blend of 1/3/6-month)

RankSector1m3m6mblend
1Technology-3.0%+39.6%+24.1%+20.2%
2Industrials+4.5%+14.1%+15.7%+11.4%
3Real Estate+2.8%+14.1%+13.6%+10.1%
4Health Care+6.7%+12.4%+3.8%+7.6%
5Materials+0.8%+5.9%+13.5%+6.7%
6Utilities+4.2%+2.0%+9.4%+5.2%
7Consumer Staples+1.0%+4.3%+9.8%+5.0%
8Financials+4.9%+12.4%-3.0%+4.8%
9Energy-4.8%-13.3%+23.0%+1.6%
10Consumer Disc-6.1%+8.4%-6.3%-1.3%
11Comm Services-8.8%-0.5%-9.6%-6.3%

Top-3 momentum rotation vs SPY (65 months)

  • Rotation final equity: 1.831x vs SPY 2.221x β†’ rotation LAGGED buy-and-hold (0.82x relative).

Read this

  • The top of the ranking is where to lean insider-cluster sizing (ties to the sector finding).
  • If rotation only matches SPY, sector timing isn't worth the turnover β€” use the ranking as a filter, not a standalone strategy.

Aggregate Insider Sentiment β€” a market-timing layer

New axis: instead of *which stock* insiders buy, this asks whether the whole market's insiders are tilting toward buying β€” and whether that predicts SPY. Monthly index from 43 months of the buy/sell panel (2016-03β†’2026-09), dollar-weighted:

sentiment = $buys / ($buys + $sells) β€” 1.0 = all buying, 0.0 = all selling.

Current reading (2026-09): 0.20 (z = -0.24 vs trailing 12m).

Dollar-weighted sentiment β†’ forward SPY return

HorizonHIGH buy-tilt (z>+1)NEUTRALLOW buy-tilt (z<βˆ’1)all months
+1mβ€” (n<12)+6.04% [CI +3.55%,+8.29%] n=21 βœ…β€” (n<12)+4.10% [CI +1.03%,+6.99%] n=30 βœ…
+3mβ€” (n<12)+14.94% [CI +9.69%,+19.84%] n=19 βœ…β€” (n<12)+12.40% [CI +6.91%,+17.71%] n=28 βœ…
+6mβ€” (n<12)+28.47% [CI +18.99%,+37.33%] n=17 βœ…β€” (n<12)+27.11% [CI +18.93%,+34.98%] n=25 βœ…

Count-weighted sentiment β†’ forward SPY return

HorizonHIGH buy-tilt (z>+1)NEUTRALLOW buy-tilt (z<βˆ’1)all months
+1mβ€” (n<12)+3.31% [CI -1.83%,+7.83%] n=16 nsβ€” (n<12)+4.10% [CI +0.97%,+6.93%] n=30 βœ…
+3mβ€” (n<12)+14.10% [CI +6.64%,+20.70%] n=14 βœ…β€” (n<12)+12.40% [CI +6.86%,+17.65%] n=28 βœ…
+6mβ€” (n<12)+32.07% [CI +22.20%,+40.63%] n=13 βœ…β€” (n<12)+27.11% [CI +18.95%,+35.03%] n=25 βœ…

Read this

  • If HIGH buy-tilt months beat LOW months at +3/+6m, aggregate insider sentiment is a genuine market-timing input β€” something the firm currently has no insider-based version of.
  • If all buckets β‰ˆ the all-months baseline, insiders predict their own stocks but not the index β€” a clean negative worth publishing.
  • Caveat: the panel is cluster-filtered (β‰₯2 insiders), so this is 'cluster sentiment', not all Form 4 activity. Monthly n is small by construction β€” treat CIs, not point estimates, as the answer.

Earnings 8-K Language β†’ Reaction Study

Scraped the press-release text of every item-2.02 8-K for 18 large-cap companies, scored each by a finance-tuned positive/negative word list, and correlated the score with the 2-session abnormal return vs SPY.

Headline

  • n = 943 earnings prints priced
  • Correlation (sentiment vs abnormal return): r = -0.007
  • High-sentiment tercile mean abn: +0.32%
  • Low-sentiment tercile mean abn: +0.68%
  • High βˆ’ low spread: -0.36%

Per-ticker correlation

TickernSentiment ↔ next-session abn
META50-0.331
KO50+0.319
JPM50+0.277
NVDA53-0.274
AVGO32+0.259
UNH50+0.256
HD50-0.206
AMZN50-0.142
V50+0.109
MSFT50+0.103
XOM61+0.086
WMT50-0.079
GOOGL44+0.062
ORCL50+0.057
AAPL51+0.019
TSLA99+0.019
BAC52+0.015
MA51-0.010

How to read this

  • A positive correlation means upbeat language predicts a positive reaction (the words leak the news).
  • A correlation near zero means the prose is just prose β€” the market reads the numbers, not the adjectives.
  • This is a simple word-list classifier; it is honest evidence about whether language alone has signal β€” not a full NLP.

🌐 Market Context & Watchlists

Live backdrop and candidate trades, refreshed by the weekly job.

Data Freshness β€” are the feeds actually alive?

Every dataset here is supposed to grow. Two of them silently stopped in 2026 while still reporting success every Sunday (the price cache fetched each ticker once and never refreshed; the sell scraper had a hardcoded date cutoff). This page checks the last date inside the data, which a no-op rerun cannot fake.

Checked 2026-10-23.

FeedLast dataAgeToleranceStatus
price cache (closes)2026-10-230d7dβœ… fresh
open cache2026-10-230d10dβœ… fresh
insider buys (raw)2026-10-230d10dβœ… fresh
insider sells (raw)2026-10-230d10dβœ… fresh
clusters (priced)2026-09-3023d45dβœ… fresh
IV snapshots2026-10-185d10dβœ… fresh
fundamentals snapshots2026-10-185d10dβœ… fresh
gov contract awards2026-12-01-39d21dβœ… fresh

⚠️ 3 of the last 6 months have buys but no sells (or vice versa) β€” the signature of one half of a paired feed dying. The sentiment index drops these months rather than reading them as 'everyone is buying'.

βœ… All critical feeds within tolerance.

Read this

  • Green pipeline logs mean the scripts ran, not that the data moved. This is the check that tells the difference.
  • It exits non-zero on a stale critical feed, so a freeze shows up as a FAILED step in the weekly log instead of hiding for months.

Cross-Asset Macro Backdrop

Rolling correlations + momentum across SPY and macro proxies (gold, oil, dollar, bonds, high-yield, bitcoin, semis, energy, financials). Context for conditioning equity signals on the macro regime.

SPY correlation β€” 3yr vs last 63d (regime shift)

Asset3yr corrrecent 63dΞ”
Energy+0.31-0.30-0.62
Oil-0.03-0.55-0.52
Long Bonds+0.15+0.39+0.24
Financials+0.72+0.52-0.21
Gold+0.18+0.37+0.19
US Dollar-0.09-0.25-0.16
Semis+0.80+0.68-0.12
High Yield+0.76+0.72-0.04
Bitcoin+0.38+0.37-0.01

63-day momentum (3-month trend)

Asset63d return
Oil+25.5%
Bitcoin+21.3%
Energy+17.3%
S&P 500+2.2%
Gold+1.9%
US Dollar+1.5%
Financials-0.7%
High Yield-2.1%
Long Bonds-3.2%
Semis-7.6%

Read this

  • Big positive Ξ” = an asset that's started moving WITH stocks lately (diversification breaking down β€” risk-on or stress).
  • Dollar (UUP) & bonds (TLT) turning positively correlated with SPY often flags a liquidity-driven 'everything' move.
  • Momentum ranking shows the macro leaders/laggards to lean equity exposure toward/against.

Market Breadth Layer β€” % of S&P 500 above its 200-day MA

Daily series, 2,578 days, computed from the local price cache. Breadth distinguishes narrow-leadership rallies from broad ones β€” a regime dimension the SPY-only gate can't see.

Current reading (2026-06-15): 63% of the index above its 200d MA.

ReadingInterpretation
> 70%broad participation β€” risk-on confirmed
40–70%mixed β€” SPY gate alone may mislead
< 40%narrow/deteriorating β€” treat BULL signals skeptically

Read this

  • Now joinable to every event study via date β€” future filters can require e.g. SPY>200MA and breadth>50% (a stricter, cleaner BULL).
  • Series regenerates weekly from cached prices at zero API cost.

Insider SELL-Side Layer β€” the untouched half

All prior insider work used BUYS only. This layer adds 44,001 cluster-sale transactions (2016β†’present, same screener) and derives a monthly buy/sell imbalance panel: 41,003 ticker-months (42,005 buy-txns vs 44,000 sell-txns).

Why this compounds

  • Enables imbalance signals the literature actually uses (buy-ratio spikes, distribution-top warnings when officers dump into strength).
  • A future study can test: does a ticker-month with buys>0 AND zero sells beat one where insiders are selling into the same window?
  • Refreshes weekly with the deep-scrape job; the panel grows every Sunday.

Status

Layer collection only β€” no signal claim yet. First imbalance study queued once a fresh quarter of joint data accumulates.

Pairs Scan v2 β€” full S&P 500, corr β‰₯ 0.75

Scanned 499 S&P 500 names (2y daily, 252d ratio window). 259 pairs cleared corr β‰₯ 0.75; 32 are dislocated at |z| β‰₯ 2.0 right now.

v1 found just 5 pairs at the 0.85 bar β€” one became the live MA/V trade. This is the fatter watchlist.

Top 40 by |z|

PairCorrzRatio now252d mean
BAC / JPM0.79-4.650.15080.1714
KMI / WMB0.86+3.950.47980.4436
PNC / USB0.90-3.683.70153.9437
MTB / USB0.88-3.443.66563.9118
AON / MRSH0.75-3.001.55321.8838
DTE / PNW0.76-2.891.25441.4298
ADP / PAYX0.86+2.892.5882.3094
AEE / DTE0.81+2.810.81770.7688
COP / DVN0.85+2.692.8882.6146
NWS / NWSA0.95-2.641.09491.136
DTE / EVRG0.78-2.601.55741.736
COP / FANG0.83+2.590.71840.6288
CCL / NCLH0.83+2.551.72091.412
DTE / LNT0.83-2.461.89891.9932
EOG / FANG0.82+2.450.77470.7139
CFG / USB0.87-2.411.07881.1336
APA / OXY0.83+2.410.78790.6427
LNT / PNW0.79-2.370.66060.7172
ACGL / HIG0.76+2.330.74950.7141
APA / FANG0.81+2.330.2410.1875
KEY / USB0.85-2.330.34040.375
NRG / VST0.78-2.310.68250.8976
VTR / WELL0.80-2.300.36560.3974
RF / USB0.87-2.300.46740.5014
FITB / USB0.88-2.270.85770.8974
MPC / PSX0.87+2.141.63071.4208
MET / PRU0.83+2.120.86930.7766
COP / OXY0.82+2.102.34872.16
APA / DVN0.83+2.090.96890.78
EVRG / LNT0.80+2.081.21931.1494
AEE / PNW0.81-2.021.02571.0986
DTE / WEC0.81-2.011.19341.2538
DTE / DUK0.75-1.991.0681.1269
CMS / ED0.78-1.990.6120.6844
ADI / MCHP0.83+1.965.35464.5622
BLDR / LEN0.76-1.950.7090.904
HIG / TRV0.81-1.940.34690.4313
NCLH / RCL0.76-1.930.05420.0689
DUK / ED0.82-1.901.0821.1615
ESS / MAA0.82+1.892.31012.0628

Read this

  • Enter at |z| β‰₯ 2 (long the cheap leg, short the rich), exit near z = 0. Same playbook as MA/V.
  • The 0.75 bar admits more spurious correlation than 0.85 β€” eyeball the pair for an economic link (same sector/business) before trading. No automated sector check here yet; that's the v3 upgrade.
  • Pairs trading decays in high-VIX regimes; gate entries on VIX < 22.

πŸ“ Archive β€” earlier & superseded studies

Kept for the record. Numbers here are superseded by the Verdict section above.

tap to expand Β· 20 studies

Top-100 13F Filer Study

Extended the 19-filer disclosure-bump study to 85 famous 13F filers, 96491 new-stake disclosures priced. Real public data: SEC EDGAR + yfinance.

Pooled

  • n = 96491 Β· Win% 49.1% Β· Mean abn +1.73% Β· p = 1.0 (not significant)

Out-of-sample

WindownWin%MeanpSig?
Train (<2022)4339048%+0.03%1.0no
Test (2022+)5310150%+3.11%0.14682no

Statistically significant filers (p < 0.05)

FilernWin%Mean abnp
Berkshire Hathaway5671%+2.04%0.00092
Eminence Capital24856%+0.92%0.03267
ARK Investment40857%+0.60%0.00431
Lone Pine Capital19756%+0.45%0.04351
Eaton Vance153653%+0.26%0.00497
Diamond Hill25857%+0.25%0.01457
Duquesne Family Office50754%+0.24%0.02529
Third Point23058%+0.23%0.0104
Norges Bank743751%+0.21%0.00613
Holocene Advisors444853%+0.21%0.00012
Point72199652%+0.16%0.04239
Maverick Capital297552%+0.16%0.03073

12 of 85 filers cleared p<0.05. Conviction managers cluster at the top; high-turnover quant shops do not β€” exactly the pattern the 19-filer study found.

Top-100 13F β€” Dissection

The Top-100 13F headline study (n=96,491) looked great in-sample at +1.7% but collapsed to +0.18% under walk-forward. This script slices the same events by who filed, who manages the money, which ticker, and which year β€” searching for a sub-population that holds up.

Best filers (β‰₯30 events)

FilernWin%Mean abn
Berkshire Hathaway5671%+2.04%
Scion Asset Mgmt13757%+0.93%
Eminence Capital24856%+0.92%
Eagle Capital6555%+0.79%
Lansdowne Partners14954%+0.71%
Generation Investment7858%+0.65%
ARK Investment40857%+0.60%
Southeastern Asset6458%+0.56%
Whale Rock29554%+0.54%
Whale Rock Capital29554%+0.54%
Pelham Capital5252%+0.50%
PRIMECAP Management24453%+0.50%
Oaktree Capital81050%+0.49%
Causeway Capital21748%+0.47%
Tiger Global20453%+0.46%

Worst filers (β‰₯30 events)

FilernWin%Mean abn
Capital World Investors102649%-0.15%
Bridgewater Associates345345%-0.18%
Pzena Investment25245%-0.19%
Cantillon Capital3444%-0.28%
Light Street23851%-0.32%
Yacktman Asset Mgmt6844%-0.43%
Aristotle Capital8138%-0.44%
Geode Capital299243%-0.54%
Appaloosa3047%-0.59%
Tybourne Capital5139%-0.83%

Best tickers (β‰₯10 stakes opened)

TickernWin%Mean abn
CYN1050%+11.78%
TRON1958%+9.86%
OPENZ2584%+9.58%
BCAB1889%+8.24%
SNNRF11100%+6.17%
QS4055%+6.09%
CRWV1953%+5.97%
PXS1765%+5.87%
WLTH1283%+5.64%
SPCE3667%+5.33%
NVTS2374%+5.28%
DCTH1547%+5.26%
ELOX1060%+4.98%
MREO1191%+4.81%
LCTX1464%+4.71%

By year β€” is the edge persistent?

YearnWin%Mean abn
20136045%+0.28%
201435149%+0.06%
201536041%-0.37%
20161,77251%+0.42%
20176,55048%-0.02%
20186,71349%-0.16%
20197,31945%-0.17%
20208,59852%+0.46%
202111,66245%-0.15%
20229,60253%+0.49%
202310,03352%+0.32%
202412,39154%+0.34%
202514,28944%-0.23%
20266,77350%-0.20%

Read this

  • The headline collapsed in walk-forward, so the question is: is there a filer or ticker sub-cohort with a real edge?
  • A filer in the top-15 with mean β‰₯+1% and nβ‰₯100 is a candidate for a dedicated bot.
  • A by-year table that shows the mean halving or flipping post-2020 is evidence the edge is fading β€” the market has caught on.
  • If every cohort lands within Β±0.5% of zero, the conclusion is that the broad Top-100 13F signal is noise. Publish the null.

S&P 500 Earnings-Reaction Study

Every item-2.02 8-K filing (2015β†’2026) for the current S&P 500. 23128 earnings reactions measured as the 2-session abnormal return vs SPY around each filing date.

Pooled across all S&P 500 prints

  • n = 23128
  • Mean absolute move: +4.37%
  • % of prints that moved >3%: 53%
  • % of prints that moved >5%: 32%
  • Mean signed drift: +0.26%
  • Biggest single up move: +48.3%
  • Biggest single down move: -47.3%

Top 25 biggest-mover tickers (by mean absolute move)

TickernMeanabn% >5%Biggest upBiggest down
CVNA3815.79%84%+40.8%-35.4%
APP2115.19%76%+46.0%-20.8%
TTD4015.18%65%+45.5%-40.2%
PLTR2314.40%91%+29.3%-20.1%
AXON4611.82%80%+30.1%-18.9%
DDOG2711.76%70%+38.8%-20.9%
SMCI6211.72%73%+30.9%-27.0%
ALGN459.84%62%+35.1%-36.8%
VRT279.83%67%+31.7%-36.4%
COHR469.62%67%+24.5%-27.4%
CIEN479.53%68%+24.4%-22.1%
LULU459.40%78%+18.7%-21.4%
XYZ429.37%71%+31.3%-24.0%
FTNT488.90%67%+26.0%-22.9%
NFLX488.87%65%+20.6%-34.6%
SNDK58.77%100%+11.6%-5.0%
ZBRA468.65%78%+16.2%-20.0%
GNRC478.65%60%+29.2%-26.1%
LITE478.57%74%+24.4%-20.0%
DECK468.51%63%+22.7%-19.0%
SW88.50%62%+17.4%-15.3%
DLTR458.41%69%+19.0%-17.9%
MRNA388.35%61%+19.3%-32.3%
AKAM468.28%67%+20.6%-21.3%
PANW458.24%58%+19.2%-28.0%

Read this

  • The S&P average earnings reaction is the base rate to compare any single-stock claim against.
  • A stock above the pooled mean β€” e.g. META's 8.2% β€” really does move more on earnings than the typical company.
  • All EDGAR + yfinance, free and reproducible.

Congressional PTR Study

Skipped β€” public PTR datasets were not reachable at scrape time.

The House Clerk and Senate eFD bulk disclosures are available as PDFs and XML at https://disclosures-clerk.house.gov/ and https://efdsearch.senate.gov/ but those formats need OCR/parsing the script does not implement. A future run should target those primary sources directly. No data, no claim.

Form 4 SP500 β€” Open-Market Insider Purchases Study

Every open-market PURCHASE (transaction code P) disclosed on Form 4 by any officer/director of an S&P 500 company since 2020. 717 priced events. Generalises the Dimon-buy finding to the full SP500 universe.

Headline

  • Any insider open-market buy β†’ +5d abnormal: n=717 Β· Win% 55.1% Β· Mean abn +0.57% Β· p=0.00356 (significant)
  • Any insider open-market buy β†’ +20d abnormal: n=643 Β· Win% 52.6% Β· Mean abn +0.18% Β· p=0.10346 (not significant)
  • Cluster buys (2+ insiders same day) β†’ +5d: n=63 Β· Win% 57.1% Β· Mean abn +1.07% Β· p=0.15675 (not significant)

Read this

  • The Dimon-buy signal was 67% / +5.5% over 6 events. Pooled across the whole S&P this number tells us whether that finding generalises.
  • Cluster buys (multiple insiders the same day) are the strongest version β€” a 'someone knows' signal.
  • All EDGAR Form 4 + yfinance, free and reproducible.

Insider Cluster Buying

When β‰₯2 distinct insiders at the same S&P 500 issuer file open-market buys within 30 days, that's a cluster. We measure the +5d and +20d abnormal return on the first event in the cluster.

63 clusters detected across the dataset.

By cluster size

# insidersn events+5d Win/Mean+20d Win/Mean
2 insiders4262% / +1.42%69% / +2.51%
3 insiders1258% / +1.74%56% / +2.11%
4 insiders367% / +2.66%0% / -5.19%
5 insiders10% / -3.26%0% / -9.00%
6+ insiders520% / -3.61%50% / +1.56%

Recent clusters (last 20)

DateTickerInsidersEvents
2026-05-14JKHY21
2026-05-13TKO31
2026-05-13ZTS21
2026-05-12ADT31
2026-05-11NCLH21
2026-05-08POOL31
2026-05-07GEHC21
2026-05-04EPAM61
2026-05-04SPGI21
2026-04-30GEHC21
2026-04-28CHTR21
2026-04-15CAG21
2026-04-14NKE21
2026-04-09TPL21
2026-04-01SPG101
2026-03-16IFF21
2026-03-13CRH21
2026-03-11SRE21
2026-03-03CSGP21
2026-03-03KKR31

Read this

  • A monotone lift across cluster sizes (2 β†’ 3 β†’ 4+) means conviction stacking works.
  • If 3+ insider clusters double the +20d mean vs single-insider Dimon baseline (+5.5%), the Insider Cluster Bot is real.
  • Recent clusters table is the live watchlist β€” these are the names to forward-track.

Post-Earnings Announcement Drift (PEAD)

Built 371 surprise+return rows from 120 S&P 500 tickers. Surprise = (actual EPS βˆ’ consensus) / |consensus|. Each row carries abnormal return at +1/+5/+20/+60 days (stock minus SPY).

Drift by surprise bucket

Bucketn+1d Win/Mean+5d+20d+60d
BIG_BEAT (>20%)5357% / +0.34%62% / +1.78%51% / -0.24%53% / +6.47%
BEAT (5–20%)12154% / +0.38%54% / +0.34%47% / +0.59%48% / +2.31%
INLINE (-5 to 5%)16452% / +0.36%52% / +0.57%43% / -0.97%34% / -3.02%
MISS (-20 to -5%)2352% / +0.52%74% / +2.93%35% / -0.61%30% / -6.05%
BIG_MISS (<-20%)1060% / +1.34%50% / -2.02%40% / -7.24%20% / -12.36%

Read this

  • A clean PEAD shows monotone drift: BIG_BEAT mean rising from +1d β†’ +60d.
  • If +60d isn't bigger than +20d, the drift faded and the trade window is shorter.
  • BIG_MISS drifting more negative = the short PEAD trade (sell weaknesses).
  • For a PEAD-bot: enter +1d after a >20% beat, exit at the horizon where mean peaks.

Bot Discovery β€” Cross-Study Pattern Mining

Meta-analysis across 128,279 unified events from 10 studies. Surfaces patterns that suggest the *next* bots to build.

1 Β· Convergence β€” when multiple studies fire the same day on the same ticker

  • 6,844 convergence events (β‰₯2 studies, same ticker, same date)
  • Win rate: 48.8% Β· Mean abn: +0.10%
  • Bot idea: *Multi-Confirm Bot* β€” only fire when β‰₯2 desks agree on the same name. Higher signal-to-noise, lower frequency.

2 Β· Sector cross-cut β€” where the edge clusters

SectornWin%Mean abn
Tech625853%+0.47%
Consumer-Cyc292251%+0.36%
Comm/Media136754%+0.34%
Health275551%+0.18%
Energy166448%+0.12%
RealEstate148951%+0.11%
Industrials286548%+0.07%
Finance230552%+0.07%
Materials134246%+0.04%
Utilities135550%-0.01%

Bot ideas: sector-tuned variants of the existing bots β€” focus capital where the pooled cross-study edge is strongest.

3 Β· Day-of-week effects

DaynWin%Mean abn
Mon2726852%+0.26%
Tue2386049%+6.75%
Wed2609148%+0.08%
Thu2442750%+0.14%
Fri2663349%+0.02%

Bot idea: *Calendar Gate* β€” restrict any bot to its best DoWs.

4 Β· Magnitude profile β€” which signals carry fat tails

SourcenMeanabn% moves >5%
meta-earnings568.17%55%
nvda-earnings596.21%42%
sp500-earn231284.37%32%
insider-buy7173.44%21%
buffett-trim1212.90%17%
buffett-13f692.07%12%
top100-13f962053.92%10%
jpm-earnings562.29%9%
super-13f78591.73%6%

Bot idea: *Fat-Tail Bot* β€” concentrate on sources with >25% >5%-move rate and small position sizing for path survival.

5 Β· Standout tickers across all studies

Top 15 most-positive (n β‰₯ 8)

TickerSourcesnWin%Mean abn
HLNCF22528%+6282.56%
LTESF1944%+55.53%
AIHS11573%+13.29%
CYN11050%+11.78%
LODE13563%+11.61%
TRON11958%+9.86%
OPENZ12584%+9.58%
ARBK2989%+8.69%
BCAB21989%+8.34%
QUBT1967%+7.47%
UAVS11656%+6.28%
SNNRF213100%+6.13%
CRWV22250%+5.93%
PXS11765%+5.87%
WLTH21385%+5.85%

Bottom 10 most-negative

TickerSourcesnWin%Mean abn
SEPN11118%-14.31%
STUB2120%-12.63%
ANY11513%-7.87%
NVVE11625%-7.68%
RSKD11242%-7.65%
HCTI11050%-7.03%
NRDY11828%-6.59%
HTFL1100%-6.39%
UPB11354%-6.25%
GETY11338%-5.69%

Bot ideas: dedicated 'always long' watchlists for the top-15 cluster, and 'fade the pop' for the bottom-10 β€” both with the standard risk template.

Summary β€” the candidate next bots

1. Multi-Confirm Bot β€” fire only when β‰₯2 desks agree on a name + date.

2. Sector-Tuned Variants β€” recolor every existing bot for its strongest sector.

3. Calendar Gate β€” restrict bots to their best day-of-week buckets.

4. Fat-Tail Bot β€” concentrate on high-magnitude sources, sized small for survival.

5. Standout Watchlist Bot β€” the top 15 cross-study winners as a permanent long list.

All derived from existing data β€” no new scraping required to validate any of these.

Walk-Forward Simulator β€” The Honest Viability Test

For every bot, this replays its events strictly forward in time. At each event the simulator only knows what would have been knowable *then* β€” prior events' outcomes, never the future. The decision gate: take the trade only if the trailing 2-year window has WR > 50% AND mean abn > 0 over β‰₯10 prior events.

In-sample vs walk-forward β€” does the edge survive?

Botn eventsIn-sample WR / meanWalk-forward WR / meann takenVerdict
buffett-13f6971% / +1.76%68% / +1.91%
31<b style='color:#22C55E'>holds (108%)</b>
buffett-trim12148% / -0.34%41% / -1.39%
27n/a
insider-buy71755% / +0.57%55% / +0.56%
707<b style='color:#22C55E'>holds (98%)</b>
nvda-earnings5954% / +1.90%never triggered
0β€”
meta-earnings5659% / +2.03%never triggered
0β€”
jpm-earnings5645% / -0.47%never triggered
0β€”
super-13f787649% / +0.10%49% / +0.11%
2631<b style='color:#22C55E'>holds (111%)</b>
top100-13f9649149% / +1.73%50% / +0.18%
28727<b style='color:#F43F5E'>collapses (10%)</b>
sp500-earn2312852% / +0.26%52% / +0.29%
19710<b style='color:#22C55E'>holds (113%)</b>

Read this carefully β€” what each verdict means

  • Holds (β‰₯80% of in-sample edge) β€” robust. A walk-forward trader, with no foreknowledge, would have captured most of the edge in real time. Strongest candidates for live deployment.
  • Decays (40–79%) β€” fragile but real. The edge existed but a real-time trader without future knowledge would have captured significantly less. Treat the in-sample number with scepticism.
  • Collapses (<40%) β€” the in-sample 'edge' was largely a curve-fit artifact of looking at the full series. A real trader would not have seen it coming.

The honest takeaway: if a bot's walk-forward number lands near zero or negative even though the in-sample number was positive, the bot's apparent edge was you (the analyst) selecting which signals to include with hindsight. Treat that as a red flag, not a strategy.

Walk-Forward β€” the two new bots

Honest out-of-sample test of the signals found in the v2 batch. A bot only 'trades' an event if its trailing in-sample mean (events strictly before that date) was positive. If the walk-forward mean holds near the full-sample mean, the edge is real; if it collapses, it was a fit.

Results

SignalFull-sample n / mean / winWalk-forward n / mean / winVerdict
insider-cluster-2plus (abn_5d)63 / +1.07% / 57%48 / +0.41% / 54%weakens
insider-cluster-2plus (abn_20d)52 / +1.70% / 60%37 / +1.56% / 59%holds
pead-big-beat (r60 long)49 / +6.69% / 53%29 / +3.33% / 55%weakens
pead-big-miss (r60 short)5 / +6.48% / 80%β€” (never traded)β€”

Read this

  • holds β†’ promote to a dormant executor like Buffett-13F.
  • collapses β†’ shelve it; it was hindsight (same fate as Top-100 13F).
  • Small WF n means the trailing gate rarely turned on β€” needs more events before trusting.

Monte Carlo Portfolio Simulation

Loaded 128,582 signals from 10 studies. Ran 50,000 simulated portfolio paths β€” each picks a random 30% of signals and runs them through the coordinator's portfolio rules (10% per trade, max 5 concurrent, 14-day hold).

Total return β€” distribution across simulations

  • 5th percentile (bad luck): -7.3%
  • 25th percentile: +11.4%
  • Median: +27.3%
  • 75th percentile: +47.1%
  • 95th percentile (lucky): +17865.4%
  • Mean: +1465.6%
  • Probability of ending positive: 89.3%
  • Probability of any loss: 10.7%

Max drawdown β€” distribution

  • Typical (25th pct): -18.9%
  • Median: -15.0%
  • Worst 5% of paths: -26.1%

End equity from $25,000 start

  • 5th pct: $23,170 Β· Median: $31,816 Β· 95th pct: $4,491,349
  • Trades per run (median): 1311

Read this honestly

  • The median is what to expect on average. The 5th percentile is what to plan for on bad luck.
  • A wide gap between 5th and 95th = high path-dependency: the same signals could win big or lose money depending on what fires when.
  • Size your account so the 5th-percentile drawdown is survivable β€” not the median one.

Ensemble Monte Carlo v2 β€” weighted vs equal

30,000 simulated paths over 31,714 signals. Compares the correlation-weighted + quarter-Kelly portfolio against the naive equal-weight / flat-10% baseline. Same governance (max 5 concurrent, 14-day hold, $25K start).

Distribution of outcomes

MetricWeighted+KellyEqual-weight baseline
Median total return+9.5%+29.4%
Mean total return+9.6%+32.0%
5th pct (bad luck)+3.3%-6.7%
95th pct (good luck)+16.4%+79.1%
Prob(positive)99.6%90.5%
Median max drawdown-2.0%-15.7%
Worst-5% max drawdown-3.4%-26.3%

Read this

  • If weighted beats baseline on both median return and worst-5% drawdown, the correlation+Kelly construction is a real improvement β€” adopt those weights in the coordinator.
  • If weighted has lower return but much smaller drawdown, it's the risk-adjusted winner.
  • The ensemble weights and Kelly fractions used are saved in the JSON for the coordinator to read.

Cost Sensitivity Sweep

Subtracts 0 / 5 / 10 / 25 / 50 basis-point round-trip costs from every trade and re-computes win rate + mean. Many backtests look great gross and die net β€” this is the test that filters those out.

Per-bot edge at each cost tier

BotnGross (0bp)5bp10bp25bp50bpVerdict @10bp
buffett-13f69+1.76%+1.71%+1.67%+1.52%+1.26%survives (94%)
buffett-trim121-0.34%-0.39%-0.44%-0.59%-0.84%no edge
nvda-earnings59+1.90%+1.85%+1.80%+1.65%+1.40%survives (95%)
meta-earnings56+2.03%+1.98%+1.93%+1.78%+1.53%survives (95%)
jpm-earnings56-0.47%-0.52%-0.57%-0.72%-0.97%no edge
super-13f7876+0.10%+0.05%-0.00%-0.15%-0.40%DIES at 10bps
top100-13f96491+1.73%+1.68%+1.62%+1.47%+1.23%survives (94%)
sp500-earn23128+0.26%+0.21%+0.16%+0.01%-0.24%erodes (62%)

Read this

  • A typical retail trade in a liquid US stock costs ~5–15 bps round-trip after spread, slippage and commission.
  • Leveraged ETFs and small caps run higher β€” 25–50 bps is realistic.
  • A bot that dies at 10 bps is not a real strategy; the apparent edge belongs to the broker, not you.
  • The bots that survive 25–50 bps are the ones that can take leveraged-ETF translation without the cost killing them.

Kelly-Optimal Position Sizing

For each bot: win probability p, average win, average loss, payoff ratio b = avgWin/avgLoss, and the Kelly fraction f* = (pΒ·b βˆ’ q)/b. The firm sizes at quarter-Kelly by policy (lower variance, robust to estimation error). f* ≀ 0 means no edge β€” do not size.

Per-bot Kelly

BotnP(win)Payoff bFull KellyΒΌ Kelly (policy)
buffett-13f (+5d)6775%3.10+66.5%+16.6%
insider-cluster-2plus (+20d)5260%1.31+28.8%+7.2%
sp500-earn (+1d)2312852%1.06+5.8%+1.5%
super-13f (+1d)787649%1.17+5.2%+1.3%
insider-cluster (+20d)64353%0.95+2.7%+0.7%

Read this

  • ΒΌ Kelly column = recommended fraction of strategy capital per trade for that bot.
  • A bot with full Kelly > 25% has a strong edge but high variance β€” quarter-Kelly tames it.
  • Negative or near-zero Kelly = the bot's losses are as big as its wins; don't allocate.
  • Compare to the flat 5% the Buffett executor currently uses β€” Kelly may justify more or less.

Multi-Confirm Bot β€” Backtest

Tests whether trading only convergence signals (β‰₯2 desks fired on the same ticker within 5 days) beats trading any single desk alone. Bot Discovery flagged 6,837 such events; this measures whether stacking confirmation pays.

Headline

  • 11,471 multi-confirm events
  • Win rate: 49.6%
  • Mean abnormal return: +0.14%
  • Range: -50.15% β†’ +45.52%

By convergence size β€” does more confirmation help?

# desks agreeingnWin%Mean abn
2 desks1107849%+0.12%
3 desks38553%+0.50%
4 desks888%+4.44%

Single-desk baselines (for comparison)

DesknWin%Mean abn
buffett-13f6971%+1.76%
insider-buy71755%+0.57%
jpm-earnings5645%-0.47%
meta-earnings5659%+2.03%
nvda-earnings5954%+1.90%
sp500-earn2312852%+0.26%
super-13f787649%+0.10%
top100-13f9649149%+1.73%

Read this

  • If the multi-confirm number beats every single-desk number, the Multi-Confirm Bot is a real upgrade β€” confirmation is signal.
  • If multi-confirm is close to the strongest single desk, you are paying frequency for marginal quality improvement.
  • If the bigger-cluster rows (3+, 4+ desks) keep getting better, the bot should require more confirmation, not less.

Pairwise Correlation & Ensemble Weights

Bucketed each bot's events into monthly mean abnormal return, then computed Pearson correlation across the 9 bots with enough history (β‰₯12 monthly obs). Bots that correlate near zero with the rest are the most valuable additions to a portfolio β€” they contribute return without adding the same risk twice.

Correlation matrix

buffett-13fbuffett-triminsider-buyjpm-earnmeta-earnnvda-earnsp500-earnsuper-13ftop100-13f
buffett-13f+1.000-0.056+0.219β€”β€”-0.223+0.081+0.199+0.091
buffett-trim-0.056+1.000+0.199β€”-0.284+0.041+0.174+0.127+0.399
insider-buy+0.219+0.199+1.000-0.150+0.066+0.017+0.104+0.197-0.153
jpm-earnβ€”β€”-0.150+1.000-0.086β€”+0.227β€”-0.053
meta-earnβ€”-0.284+0.066-0.086+1.000+0.100+0.113+0.060+0.231
nvda-earn-0.223+0.041+0.017β€”+0.100+1.000+0.039-0.060+0.099
sp500-earn+0.081+0.174+0.104+0.227+0.113+0.039+1.000+0.050+0.062
super-13f+0.199+0.127+0.197β€”+0.060-0.060+0.050+1.000+0.429
top100-13f+0.091+0.399-0.153-0.053+0.231+0.099+0.062+0.429+1.000

Proposed ensemble weights

Weights penalize bots that overlap with others and reward standalone positive mean.

BotMonthly mean abnAvgcorrWeight
buffett-13f+1.84%0.1424.5%
buffett-trim-0.44%0.180.0%
insider-buy+0.78%0.1410.5%
jpm-earn-0.47%0.130.0%
meta-earn+2.03%0.1327.5%
nvda-earn+2.05%0.0830.3%
sp500-earn+0.21%0.113.0%
super-13f+0.24%0.163.1%
top100-13f+0.09%0.191.2%

Read this

  • High weight = decorrelated AND positive. The highest-weighted bots are the ones to lean into.
  • Zero weight = either negative mean (no edge) or fully redundant with another bot.
  • Look for pairs with |corr| < 0.2 β€” those are independent return streams worth combining.
  • High pairwise correlation between two filings bots (e.g. Buffett vs super-13F) means they're restating the same insider-tracking edge, not stacking it.

Parameter Sweep β€” the survivors

Tuning two bots that already survived walk-forward + cost sensitivity: Buffett 13F new stakes and Dimon Form 4 open-market buys. The goal is to find the version of each signal that earns the highest mean β€” and to flag where slicing is just curve-fitting on small N.

Buffett 13F new stake

Total events: 69.

Horizon sweep

HorizonnWin%MeanMedian
abn_1d6971%+1.76%+1.02%
abn_3d6981%+2.74%+1.74%
abn_5d6775%+3.01%+1.69%

Best horizon: abn_5d.

Quality filter β€” Position value at filing (USD)

Slicing by value_usd at horizon abn_5d.

QuartilenWin%Mean
Q1 (smallest)6775%+3.01%
Q26775%+3.01%
Q36775%+3.01%
Q4 (largest)6775%+3.01%

Insider-cluster buy

Total events: 717.

Horizon sweep

HorizonnWin%MeanMedian
abn_5d71755%+0.57%+0.46%
abn_20d64353%+0.18%+0.26%

Best horizon: abn_5d.

Quality filter β€” Number of insiders buying

Slicing by n_buyers at horizon abn_5d.

QuartilenWin%Mean
Q1 (smallest)65455%+0.53%
Q265455%+0.53%
Q365455%+0.53%
Q4 (largest)65455%+0.53%

Read this

  • A monotone trend across quartiles (e.g. bigger buys β†’ bigger returns) is a real signal worth filtering on.
  • A U-shape or noise across quartiles means the size filter is curve-fitting; ignore it.
  • A horizon that beats the others by <0.5% with n<100 is not significant β€” pick the longer horizon (more events) instead.
  • Use this to set the executor's hold period and any minimum-conviction filter.

Regime Gating Sweep

Each bot's mean abnormal return, recomputed under four market regimes: SPY above/below its 200-day moving average, and VIX below/above 20. Bots whose performance flips by regime get a regime gate in the executor β€” only fire when the regime is right.

Per-bot by regime

BotAllSPY>200dSPY<200dVIX<20VIXβ‰₯20Gate verdict
buffett-13f+1.76% (n=69)+1.62% (n=56)+2.39% (n=13)+1.53% (n=47)+2.27% (n=22)fire only SPY<200d
buffett-trim-0.34% (n=121)-0.65% (n=104)+1.55% (n=17)-0.65% (n=98)+1.00% (n=23)fire only SPY<200d
insider-buy+0.57% (n=717)+0.63% (n=627)+0.18% (n=89)+0.58% (n=478)+0.56% (n=238)no clear gate
super-13f+0.10% (n=7876)+0.06% (n=6651)+0.28% (n=1225)+0.08% (n=5578)+0.12% (n=2298)no clear gate
top100-13f+0.08% (n=96473)+0.02% (n=82448)+0.44% (n=14025)-0.00% (n=70508)+0.32% (n=25965)no clear gate
sp500-earn+0.26% (n=23128)+0.23% (n=18641)+0.37% (n=4481)+0.27% (n=16319)+0.25% (n=6803)no clear gate

Read this

  • Bots that look the same in every regime are robust β€” leave them un-gated.
  • A bot that's <b>strongly positive in one regime and ~zero in the other</b> should be gated. Halves the trade count, lifts the win rate.
  • Bots that go <b>negative in one regime</b> are the most important to gate β€” they're contributing losses for free in the wrong environment.
  • Note: regimes are not independent (SPY<200d typically pairs with high VIX). The two gates often produce similar filters.

Berkshire Exit / Trim Study

Honest stratification of 121 Berkshire trim/exit/add events. Long Buffett-13F (new stakes) wins at +5d / 75% / +3.01%. Question for this study: do trims/exits *underperform* enough to make a short or 'avoid' bot?

By event type

Event typen (+5d)+1d Win/Mean+3d+5d
unknown12152% / +0.29%β€”48% / -0.34%

Overall

HorizonnWin %MeanMedian
abn_1d12152%+0.29%+0.04%
abn_5d12148%-0.34%-0.09%

Read this

  • Negative mean on exit or trim at +5d β†’ real short signal. Pair with Buffett-13F long.
  • If mean is positive on trims, Berkshire is selling *into* strength and the move continues. No short.
  • Use this to decide whether the Buffett executor should also short on trims.

Pairs Scan β€” mean-reversion candidates

Scanned 107 large-caps, kept pairs with |return correlation| β‰₯ 0.85 over the last 2 years (n=5 qualifying pairs). For each, computed the trailing-252d z-score of the price ratio. Pairs at |z| β‰₯ 2 are statistically dislocated β€” candidate mean-reversion trades.

Top 25 by |z|

PairCorrRatio nowMean ratioz
MA / V0.871.51011.6343-2.71
GOOG / GOOGL1.000.98961.0011-2.53
GS / MS0.884.92925.0048-0.51
HD / LOW0.881.46431.5042-0.38
PNC / USB0.873.99874.028-0.24

Read this

  • A pair at z = +2 means the ratio is 2 standard deviations *above* its 1y mean. Long the cheap leg, short the rich one.
  • The bot enters at |z| β‰₯ 2 and exits at |z| ≀ 0.5.
  • Pairs trading works best in low-vol regimes; gate on VIX < 22.
  • High correlation is necessary but not sufficient β€” confirm sector match before sizing real capital.
Reproducible research Β· free public data (SEC EDGAR Β· yfinance Β· public PTR datasets) Β· educational use only Β· Back to Premium