Think a director selling nearly a billion dollars of stock is a sell signal? That instinct is exactly the kind of thing that turns a promising trading bot into an expensive lesson. I build bots for a living, mostly data-ingestion pipelines and alerting systems, and insider filings are one of the most requested data sources I get asked about. They are also one of the easiest to get wrong.
Here is what happened in the third quarter of 2026. Mark Stevens, a director at Nvidia, topped the list of insider sellers with close to $947 million in Nvidia stock. Jayshree Ullal followed with roughly $446 million in Arista Networks stock. Jeff Bezos came third among the AI-adjacent names, offloading over 1.2 million Amazon shares for about $346.5 million. Between the three of them sits a combined fortune of $302 billion.
Stevens was not making a one-off move either. He sold $38.5 million in Nvidia shares in March 2026, then $186 million in June 2026. Add the third-quarter sales and you get more than $534 million moved in under a year, with filings in early September 2026 covering 1,022,239 shares sold through trusts holding his indirect interests.
Why the naive bot gets this wrong
The first version of an insider-tracking bot almost always looks like this: poll a filings feed, parse the transaction code, compare the dollar amount against a threshold, fire an alert. Ship it, watch it scream.
That design treats every sale as an opinion about the company. It is not. A sale can be a scheduled liquidity event, a trust restructuring, a tax payment, a diversification move after years of appreciation, or an actual change of conviction. The filing format tells you the mechanics, not the motive. Your bot cannot read minds, and if you build it as though it can, you have shipped a sentiment engine dressed up as a data product.
Notice how much context the Stevens numbers need before they mean anything. Multiple sales across multiple quarters. Indirect holdings through trusts. A long-standing early position in the company. Strip that away and you are left with a big number and a scary verb.
What I would actually build
If I were standing this up today, I would separate the system into layers and resist the urge to put judgment anywhere near the ingestion path.
- Ingestion — pull raw filings on a schedule, store them immutably, and never transform on write. You want the original document for every claim your bot later makes.
- Normalization — resolve the messy parts: direct versus indirect ownership, trust and entity names that map back to one person, multiple transactions filed across adjacent dates that are really one event.
- Aggregation — roll up by person and by rolling window. One $186 million sale and three smaller ones in the same quarter should surface as a pattern, not four disconnected pings.
- Context enrichment — attach the filer’s role, holdings history, and prior sales so a human sees the trend line instead of a single data point.
- Presentation — report what happened, link the source, and stop there.
That last layer is where most builders overreach. The temptation is to slap a confidence score on the output and call it an insight. Resist it. A bot that says “director sold X shares on date Y, here is the filing, here are their last four transactions” is genuinely useful. A bot that says “bearish signal detected” is making a claim it cannot support.
The entity resolution problem nobody warns you about
If you take one practical thing from this, make it this. Sales routed through trusts and holding entities are the single biggest source of broken aggregation in these pipelines. The same individual can appear under several entity names in the same week. If your bot keys on the filer string, your totals will be wrong, and they will be wrong in the direction of understating activity. Build a mapping table, maintain it by hand if you have to, and write tests against known multi-entity filings.
What this data is good for
These numbers are a useful reminder that the AI boom is producing real, realized wealth, and that the people closest to it are converting paper gains into cash on a schedule. That is an observation about market structure, not a prediction. Your bot can measure the first thing precisely. It cannot measure the second at all.
Build the measuring tool. Keep the interpretation where it belongs, with the person reading the alert. The craft in bot building is not in making the machine louder, it is in making it honest about what it actually knows.
🕒 Published: