A single insider buy shows up in the feed and it barely moves the needle for me anymore. Someone bought $80k of their own stock. Maybe they believe in the company. Maybe their comp advisor told them the optics look good. Maybe they just got a bonus and parked it somewhere familiar. You genuinely cannot tell from one print, and pretending you can is how people talk themselves into bad trades.
What actually gets my attention is when three or four of them do it in the same week. The CEO buys, then the CFO buys two days later, then a couple of directors follow inside the same window. Now the innocent stories have to line up in a way that gets less and less likely with each new buyer. One person needing a down payment on a house is plausible. Four insiders independently needing liquidity events that resolve into buying the same stock at the same time is not really a coincidence you can wave away.
Why the cluster beats the lone buy
The mechanism is basically statistical. Any individual insider has personal reasons to buy or sell that have nothing to do with what they know about the business. Diversification, taxes, divorce, a kid's tuition, an exercise-and-hold on options. Academics call this the noise, and there is a lot of it. The classic Cohen, Malloy and Pomorski work made the useful distinction between routine insiders, the ones who trade on a predictable calendar every year, and opportunistic insiders, the ones whose trades are irregular and actually predict returns. The routine buys are mostly noise. The opportunistic ones carry information.
A cluster is a shortcut to that same idea without having to model each person's history. When several insiders act together in a tight window, the odiosyncratic reasons cancel out and what is left is the thing they have in common, which is a view on the company. It is the difference between one person leaving a party early and half the room heading for the door at once. The second one tells you something changed.
There is also a softer human layer to it. Insiders talk. A board that just sat through a strong quarterly review, or got early read on a contract win, or watched the stock get punished for something they think is temporary, will sometimes move as a loose group even without coordinating. The buying itself is a form of shared conviction leaking out into the filings.
Sizing it so a big number does not fool you
The raw dollar figure is the first thing everyone looks at and the first thing that misleads you. A director buying $500k sounds huge until you find out they are worth $200 million and sit on six other boards. That is a rounding error for them. Meanwhile a CFO on a $600k salary putting $250k of after-tax money into the stock is making a real bet, because that is close to a full year of take-home going into one name.
So I try to scale every buy against the person, not against the market cap. A few things I weight:
- Dollar size relative to salary. A buy that is a meaningful fraction of annual cash comp means more than a bigger absolute buy from someone for whom it is pocket change.
- Change in stake, not just the dollar amount. Someone increasing their holdings by 40 percent is louder than someone adding 2 percent to an already large position.
- Open-market purchases only. Option exercises, gifts, and shares acquired through comp plans are not the same as walking into the market and paying cash. The filing codes tell you which is which, and a lot of naive detectors count them all the same and get flooded with junk.
- Role. A CEO and CFO buy is worth more than two directors, because they see the actual numbers first and have the most to lose reputationally if they are wrong.
When you stack those together, a modest-looking cluster can outrank a flashy single buy. Three people each committing a real slice of their own money beats one director writing a check that does not change their life.
The false positives that will burn you
Clusters are not magic, and there are a few specific setups where a cluster forms for reasons that have nothing to do with an information edge. If you do not filter these out, your detector will look great in backtests and then feed you garbage live.
The comp-driven window is the big one. Plenty of companies have a fixed trading window after earnings when insiders are allowed to transact, and some have programs that nudge or require executives to build ownership over time. When a bunch of buys land in the same week simply because that is the only week the legal team lets them trade, the timing is an artifact of policy, not a view on the stock. You can catch a lot of these by checking whether this same group buys in the same calendar window every year. If they do, it is routine, and you weight it way down.
The other one is the optics play. A company announces a buyback, the stock is limping, and suddenly a couple of executives make visible open-market purchases. Sometimes that is genuine conviction. Sometimes it is theater meant to signal confidence to the market while the real story underneath is weaker than the gesture. These are harder to filter mechanically, so I lean on context. Is the buying happening quietly across several people, or is it one loud purchase timed right next to a press release. Quiet and distributed reads as more honest than loud and well-timed.
How I would actually build the detector
If you were building this from scratch, the skeleton is not complicated. You are ingesting the filings, normalizing them, and looking for density in a short window with the noise stripped out.
- Pull the raw filings and keep only real open-market buys. Filter by transaction code so you drop the option exercises and comp-plan acquisitions. This step alone removes most of the false signal.
- Group by company and by a rolling window, something like seven to fifteen trading days. Count distinct insiders, not distinct transactions, so one person buying five times does not fake a cluster.
- Tag each insider routine or opportunistic using their own trade history. If they buy in roughly the same window every year, mark them routine and discount them heavily.
- Score each buy on the personal-scale metrics, dollar-to-salary, percent change in stake, and role weight, then sum across the cluster.
- Layer the context on top. Flag whether a buyback or earnings release sits inside the window, and whether the sector is in a spot where insiders tend to know something the tape does not yet.
Sector and timing matter more than people expect here. Insider buying in a small-cap biotech ahead of a data readout is a different animal from insider buying in a sleepy utility, because the information asymmetry between the boardroom and the market is far larger in the first case. And a cluster that forms after the stock has already been cut in half tends to age better than one that forms at the highs, because it is more likely to be people stepping in on something they think is mispriced rather than chasing.
At Blockcircle the insider feed is one input among several, and I mostly use a cluster as a reason to go look closer rather than as a standalone buy trigger. It tells you where the people with the best seats just voted with their own cash. That is worth a serious read, but it is a starting point for the work, not a substitute for it.