Insider— how it works

When a company's officers and directors buy their own stock on the open market, they are putting personal capital behind non-public conviction — a documented, decades-old anomaly (Lakonishok-Lee; Cohen-Malloy-Pomorski). It is an information edge, distinct from the price-momentum signals the other books trade — which is exactly why it diversifies them.

Every SEC Form-4 open-market purchase is scraped daily from OpenInsider. The universe is any stock with an insider purchase in the trailing 30 days plus tradeable liquidity. A 5-seed XGBoost ensemble then ranks that universe on 36 features:

• Conviction — role-weighted dollar value (CEO/CFO buys weighted above directors), largest single buy, total 30/90-day buy value
• Cluster quality — count of distinct insiders buying within 7 and 30 days
• Opportunistic vs routine — a causal prior-purchase count per insider; rare/first-time buyers carry more information than habitual filers
• Context — company fundamentals (solvency, dilution, profitability), price/volatility technicals, and market regime

All insider features are keyed on the Form-4 filing date (the public disclosure date) with a one-day safety lag — no lookahead.

Path-aware, asymmetric: trained on whether a stock reaches +25% within 60 days without dropping -15% the day after entry. The immediate-after-entry whipsaw is the one downside the model can reliably screen.

Daily entries: top-3 by model conviction, subject to a 7-day per-ticker reentry cap and no overlap with currently-held names. 60-day hold with four equal scale-out slices at days 15/30/45/60. Long-only, no stops — the partials are the exit. Flat position sizing.

• Sharpe 1.58 — measured on honest true-daily marks
• CAGR +57.5%, max drawdown -24%
• All 5 out-of-sample folds positive
• ~0.55 correlated to the precision25 book — a partial diversifier, not market-neutral (it still carries equity beta)

The backtest window (2024-01 → 2026-05) is a predominantly rising market — the Sharpe would be lower in a sustained downturn. Honest expectation, not a best-case slice.

Insider transactions are re-scraped from OpenInsider daily before the morning scan, so picks always act on fresh Form-4 filings. Weekly model retrain on a rolling 30-month window.