APEX factor · Insider
Insider activity treats SEC Form 4 filings as the strongest informed-trader signal available to the public. CEOs and CFOs buying with their own money — especially in clusters of multiple insiders within a short window — predict positive returns. Clustered selling by the same group predicts negative returns.
A CEO's stock is the worst-diversified position in their personal balance sheet. They already get RSU grants every year. Buying more — with cash — when they're months ahead of the market on internal forecasts is a strong directional bet. The signal degrades when (a) it's a single token purchase from a single low-rank insider, (b) it's a 10b5-1 plan filed months earlier, (c) the dollar amount is small relative to the insider's net worth. The factor weights for all three: role, cluster size, dollar significance, and plan-vs-discretionary.
for each ticker, last 90 days: buys_score = Σ insider_buy · role_weight · ($size / market_cap) · (1 - plan_flag) sells_score = Σ insider_sell · role_weight · ($size / market_cap) · (1 - plan_flag) cluster = count_distinct_insiders_30d cluster_mult = 1 + log(1 + cluster) insider = z_score( (buys_score - sells_score) · cluster_mult )
Role weight is highest for CEO/CFO, lower for directors, lowest for 10% holders (often passive). Plan flag is parsed from the Form 4 footnote — a 10b5-1 trade signals nothing about timing. Dollar size is normalised by market cap so a $5m purchase in a $1bn small-cap counts more than $5m in a $1tn mega-cap. The cluster multiplier captures Cohen-Malloy-Pomorski's finding that coordinated insider activity is far more informative than isolated trades.
Form 4 filings come from SEC EDGAR XBRL within hours of acceptance. The insider cron runs Mondays 12:15 UTC for the weekly aggregation, plus an inline refresh during the 06:00 UTC daily update for any tickers with new filings. Role classification (CEO / CFO / Director / 10% holder) is parsed from the relationship XML element. Plan-trade flag comes from the rule10b5_1Indicator field. Trades older than 90 days drop out of the calculation — the predictive content is concentrated in the first quarter after filing.
Insider clustering interacts cleanly with NLP — when officers are buying AND the latest 10-K MD&A reads optimistic (low Loughran-McDonald negative tone), the INSIDER + NARRATIVE confluence pattern fires bullish. The opposite — clustered selling plus pessimistic MD&A tone — fires the INSIDER + NARRATIVE bearish pattern. Insider also reinforces PEAD: a beat plus C-suite buying within 30 days of the announcement is a high-conviction setup.
Three honest limitations. (1) The signal is sample-limited for low-coverage tickers — many small-cap names see one Form 4 per year, so the cluster multiplier never fires. (2) Tax-driven selling masquerades as bearish — a CEO selling 10% of their stake every year on the same date is portfolio rebalancing, not a forecast, and our 10b5-1 detection is imperfect on plans filed before 2023 (the SEC tightened plan disclosure rules then). (3) The signal lags during M&A speculation — insiders trade restricted ahead of deal announcements, so we see no signal precisely when the stock is about to move.
Every ticker page shows the per-factor decomposition. The Insider score is one of twelve composing the 0–100 APEX composite.