12 breakouts, one method: the anomaly model behind Launch Sentinel's alerts

Twelve is a small number against a corpus of 79,573 products. That's by design. The model behind Launch Sentinel's breakout alerts doesn't rank the market or flag anything that clears some universal bar. It notices when one specific product breaks away from its own trajectory, and it stays quiet when it doesn't. Most anomaly detection in this corner of the industry still runs on thresholds. This one runs on history.

The threshold trap

A threshold is the obvious first build. Pick a signal score, 70 say, and flag anything that crosses it. It's simple to explain and simple to ship, and it's wrong in two directions at once. A product with a naturally high baseline sits close to 70 on an ordinary week, so routine noise trips the alert. A product with a naturally low baseline can double its signal score, break out for real, and still land nowhere near 70. Run that bar across 18 categories and 79,573 products and it mostly measures which products started closer to the line, not which ones actually changed.

Building the band

The fix is to stop comparing every product to the same number and compare each one only to itself. Every week, Launch Sentinel keeps an eight-week trailing view of a product's signal score and derives two things from it: a center (the recent normal) and a width, sized to how much that product's own score naturally moves week to week. A volatile product gets a wide band. A quiet, stable one gets a narrow band. Neither is measured against the other's history, only its own.

Eight weeks is long enough to smooth out a bad data day and short enough that a product's shift in behavior still shows up as a shift, rather than getting averaged away into an ever-fatter band that nothing could ever clear. Shorten the window and ordinary noise starts registering as a break. Lengthen it and a real change takes months to separate from the product's old normal, by which point the operator reading the alert has already missed the useful part of the window.

The hold requirement

Clearing the top of the band once is not the same as breaking out. A product can land a single strong review or ride one good week of traffic, and either fades by the following Monday. So can a single day of press coverage. If the model treated one high reading as a confirmed break, a large share of the corpus would qualify by any given Friday. So a break only counts once it holds across more than one consecutive weekly reading. The difference between a spike and a hold shows up in what happens after the peak, not at the peak itself.

Why 12 beats 200

Loosen the hold requirement, or narrow the band, and the alert count climbs fast. Run the same corpus through a version of the model that fires on a single high reading instead of a confirmed hold, and the count stops being 12. It moves closer to 200 a week, most of which drift back inside their own range within days and never should have counted at all. An operator who has to check two hundred alerts a week to find the twelve real ones is just doing manual triage.

A threshold tells you a number got big. A trajectory break tells you something changed.

Twelve confirmed breakouts across 79,573 products in 18 categories works out to roughly one in 6,600, a rate low enough that every one of the twelve is worth an operator's attention, and none of them will turn out to be a scoring quirk once someone checks. That ratio is the entire design argument: strict enough that every flagged breakout is worth a look.

What the alert actually carries

When the hold window closes and a break is confirmed, the alert carries the shape of the trajectory, not a single number crossing a line. It shows how many weeks the break has held and where the product sits inside its category of 18, plus whether press velocity moved first or lagged behind the score. Reading one of the twelve live right now looks less like a metric flashing red and more like a small case file. It shows what changed and when it started holding, measured against that exact product's own category baseline.

That's the whole design argument in one line: measure a product against itself, wait for the break to hold, and let the number of alerts fall out of that discipline instead of setting it as a target.