How Stockful's Anomaly Detection Actually Works

Plenty of inventory tools advertise “AI-powered anomaly detection” and then tell you nothing about what that actually means. That is a problem, because a black box you cannot reason about is not much use for inventory decisions. So here is exactly how Stockful’s anomaly detection works, including where it can be wrong, because the honesty is the point.

The problem it solves

You cannot watch every SKU every day. With a few hundred products across a couple of locations, the unusual things, a sudden spike, a quiet slide toward zero, get lost in the noise until they show up in a monthly report. By then the spike is a missed restock and the slide is dead stock. Anomaly detection exists to tell you when something changed, on the day it changed, not weeks later.

What counts as an anomaly

Every day, Stockful scans each SKU’s recent sales velocity, per location, and compares it against that product’s own normal pattern. It flags movement in two directions: unexpected spikes, where a product is suddenly selling much faster than usual, and drops, where it is selling much slower than usual or has stocked out against expectations.

The key word is “own.” Unusual is measured relative to each SKU’s history, not a global rule. A product that normally sells two units a day and one that sells two hundred are each judged against themselves, so a meaningful change is caught at both ends of the catalogue.

Severity, so you are not drowning in noise

Not every wobble deserves your attention, so each anomaly gets a severity tier: high, medium, or low. High-severity events can optionally trigger a Slack alert, while the rest wait quietly on the dashboard. The aim is to surface the handful that genuinely matter rather than ping you about every minor fluctuation, because an alert system you learn to ignore is worse than none.

The hypothesis, which is the part people ask about

Each anomaly comes with a one-sentence, plain-English likely cause, generated from the data around it. For example: “Velocity for Winter Jacket (Black, M) dropped 65% over the past 7 days, likely due to seasonal demand decline, the same pattern observed last year.”

It forms that hypothesis by looking at adjacent signals it can actually see: prior-year seasonality for the same product, a related variant that stocked out and may be driving substitution, a recent tag change that suggests a promotion, and so on. It does the correlation work so you do not start every investigation from a blank page.

Where it is honest about its limits

This is the important part. The hypothesis is a starting point, not a verdict, and it is grounded only in the data Stockful can see. Sometimes the real cause is external, a competitor selling out, a press mention, a spell of weather, none of which ever touches your Shopify data. In those cases the hypothesis will be incomplete, and that is by design.

We would rather hand you a useful, checkable guess than a confident wrong answer dressed up as certainty. A flag that says “here is what changed, and here is a plausible why you can verify in a minute” is something you can act on. A score with no reasoning is not.

Where to see it

Anomalies live on a dedicated page, filterable by location, severity, and direction, so you can scan the day’s flags quickly. And when a specific product has an open anomaly, a banner appears on its detail page, so you see it in context the moment you are looking at that SKU.

Why we built it this way

Inventory decisions need explanations you can reason about, not a black-box verdict. A flag with a plausible, checkable cause lets you act in seconds and learn what is going on in your own store. That is the whole design goal: not to be mysterious and clever, but to be transparent and useful.

Get started free at stockful.app. Anomaly detection is included on every plan, with a plain-English hypothesis on every flag.

More from Stockful: