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 · Win% 49.1% · Mean abn +1.73% · p = 1.0 (not significant)

Out-of-sample

Statistically significant filers (p < 0.05)

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

S&P 500 Earnings-Reaction Study

Every item-2.02 8-K filing (2015→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)

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 — e.g. META's 8.2% — really does move more on earnings than the typical company.
• All EDGAR + yfinance, free and reproducible.

Earnings 8-K Language → 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 − low spread: -0.36%

Per-ticker correlation

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 — 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 — not a full NLP.

Congressional PTR Study

Skipped — 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 — 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 → +5d abnormal: n=717 · Win% 55.1% · Mean abn +0.57% · p=0.00356 (significant)
• Any insider open-market buy → +20d abnormal: n=643 · Win% 52.6% · Mean abn +0.18% · p=0.10346 (not significant)
• Cluster buys (2+ insiders same day) → +5d: n=63 · Win% 57.1% · Mean abn +1.07% · 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 — a 'someone knows' signal.
• All EDGAR Form 4 + yfinance, free and reproducible.

Bot Discovery — Cross-Study Pattern Mining

Meta-analysis across 128,457 unified events from 10 studies. Surfaces patterns that suggest the *next* bots to build.

1 · Convergence — when multiple studies fire the same day on the same ticker

• 6,837 convergence events (≥2 studies, same ticker, same date)
• Win rate: 48.7% · Mean abn: +0.10%
• Bot idea: *Multi-Confirm Bot* — only fire when ≥2 desks agree on the same name. Higher signal-to-noise, lower frequency.

2 · Sector cross-cut — where the edge clusters

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

3 · Day-of-week effects

Bot idea: *Calendar Gate* — restrict any bot to its best DoWs.

4 · Magnitude profile — which signals carry fat tails

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

5 · Standout tickers across all studies

Top 15 most-positive (n ≥ 8)

Bottom 10 most-negative

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

Summary — the candidate next bots

1. Multi-Confirm Bot — fire only when ≥2 desks agree on a name + date.

2. Sector-Tuned Variants — recolor every existing bot for its strongest sector.

3. Calendar Gate — restrict bots to their best day-of-week buckets.

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

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

All derived from existing data — no new scraping required to validate any of these.

Walk-Forward Simulator — 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* — 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 ≥10 prior events.

In-sample vs walk-forward — does the edge survive?

Read this carefully — what each verdict means

• Holds (≥80% of in-sample edge) — 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%) — 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%) — 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.

Monte Carlo Portfolio Simulation

Loaded 128,760 signals from 10 studies. Ran 50,000 simulated portfolio paths — 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 — distribution across simulations

• 5th percentile (bad luck): -12.0%
• 25th percentile: +6.0%
• Median: +21.2%
• 75th percentile: +40.4%
• 95th percentile (lucky): +16889.4%
• Mean: +1363.8%
• Probability of ending positive: 83.5%
• Probability of any loss: 16.5%

Max drawdown — distribution

• Typical (25th pct): -21.1%
• Median: -16.5%
• Worst 5% of paths: -29.1%

End equity from $25,000 start

• 5th pct: $21,991 · Median: $30,299 · 95th pct: $4,247,358

• Trades per run (median): 1328

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 — not the median one.

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 — this is the test that filters those out.

Per-bot edge at each cost tier

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 — 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.

Multi-Confirm Bot — Backtest

Tests whether trading only convergence signals (≥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

• 11,463 multi-confirm events
• Win rate: 49.7%
• Mean abnormal return: +0.17%
• Range: -50.15% → +45.52%

By convergence size — does more confirmation help?

Single-desk baselines (for comparison)

Read this

• If the multi-confirm number beats every single-desk number, the Multi-Confirm Bot is a real upgrade — 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 (≥12 monthly obs). Bots that correlate near zero with the rest are the most valuable additions to a portfolio — they contribute return without adding the same risk twice.

Correlation matrix

Proposed ensemble weights

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

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 — 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 — 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 — and to flag where slicing is just curve-fitting on small N.

Buffett 13F new stake

Total events: 69.

Horizon sweep

Best horizon: abn_5d.

Quality filter — Position value at filing (USD)

Slicing by value_usd at horizon abn_5d.

Dimon Form 4 open-mkt buy

Total events: 895.

Horizon sweep

Best horizon: abn_20d.

Quality filter — Shares purchased

Slicing by shares_bought at horizon abn_20d.

Read this

• A monotone trend across quartiles (e.g. bigger buys → 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 — 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 — only fire when the regime is right.

Per-bot by regime

Read this

• Bots that look the same in every regime are robust — 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 — 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.

Top-100 13F — 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 — searching for a sub-population that holds up.

Best filers (≥30 events)

Worst filers (≥30 events)

Best tickers (≥10 stakes opened)

By year — is the edge persistent?

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 ≥+1% and n≥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 — the market has caught on.
• If every cohort lands within ±0.5% of zero, the conclusion is that the broad Top-100 13F signal is noise. Publish the null.