Anomaly Detection

Anomaly detection is the automated identification of values or patterns that fall outside what a dataset would normally be expected to contain. In pricing and digital shelf data it serves two purposes that get treated as one. The first is data quality: a price that arrives as 4.99 when the product has sold at 499 for a year is almost certainly a collection fault, and catching it before it reaches a report prevents a wrong number driving a real decision. The second is commercial signal: a competitor dropping 30 percent overnight is not an error but a market event that someone needs to see the same day. Both rely on knowing what normal looks like for that product, which requires enough history to establish a baseline. Detection systems applied to new listings or thin categories tend to produce noise until that history exists.

Why it matters

• Separates collection faults from market events, which are the two very different things an unexpected value can mean
• Makes monitoring scale, since nobody reviews a hundred thousand daily price observations by hand and sampling misses the outliers by design
• Determines whether alerting gets trusted, because a system that cries wolf gets muted and then the real event goes unseen

How it is used

• Price movement detection against a product's own history and its category baseline
• Data validation on collected records, catching values that parse correctly but fall outside plausible bounds
• Availability pattern detection, flagging stock behaviour that departs from a product's normal rhythm

Anomaly detection and threshold alerting solve overlapping problems differently. A threshold is a fixed rule: tell me when this price drops below twenty. Anomaly detection is relative to learned behaviour, so it flags a five percent move on a product that never moves while ignoring the same move on one that swings daily. Threshold rules are predictable and easy to explain, which matters for compliance work where a MAP breach is defined by a number rather than by unusualness. Most working setups run both.