Most of a Form 4 feed is noise

Across 103,733 non-derivative Form 4 rows in a single measured quarter, only 5.7% were open-market purchases. The rest of the volume is administrative: 26.0% tax withholding (code F), 23.8% stock grants (A), 15.7% option exercises (M), plus gifts, conversions and transfers. None of that is a person choosing to buy or sell. A tracker that shows you all of it is showing you payroll, not conviction.

The sell-side filter is the edge

Insider selling has a reputation for being uninformative, and for the most part that reputation is earned. The reason is 10b5-1: an executive adopts a written plan months in advance that sells a fixed amount on a fixed schedule, then the sale prints as a Form 4 the day it executes. It looks like a decision. It is a calendar entry. 49% of every sell row across this dataset is one of these, rising to 56% in the most recent measured quarter. They are removed before scoring. What survives is an insider who chose, on a specific day, to sell stock they were under no obligation to sell.

What the score is made of

• Size, log-scaled with a floor. Half of all insider purchases are under $15,661. Those are optics. A floor keeps them from registering, and the log scale keeps one enormous buy from swamping a page.
• Change in the insider's own stake. A $200k buy from someone who already owns $80m of stock is a rounding error. The same $200k from someone doubling their position is the actual signal. Percentage of stake is the honest measure.
• Role weight. CEOs and CFOs see the whole business. Directors see a board pack once a quarter. Ten-percent holders are usually funds and PE vehicles rather than humans, and are weighted lowest.
• Novelty. A first purchase in twelve months from a person who has never bought counts for more than the twentieth buy from a habitual accumulator.
• Cluster. Several distinct insiders moving the same way inside a short window is the single most durable pattern in this dataset. One buyer is an opinion. Four is a meeting.

Guards against bad data

Filed prices are sometimes garbage. In one measured quarter, three rows carried 99.9999% of the total dollar value of all purchases. One filer reported a price of $24,035,774 per share. Any dollar-weighted ranking that does not defend against this will put a typo at the top of the page. Prices that fall outside a sane band against the market close are flagged and excluded from dollar-weighted scoring rather than silently trusted.

What this is not

This is a filter and a score over public filings. It does not know anything the market does not already have access to, it has no view on your situation, and a tier is not a price target. The backtest tab shows what the scoring has actually done historically, including when that is unflattering.