See other articles
Anomaly Detection

Anomaly Algorithm

Early detection of anomalies plays a key role in ensuring that issues (negative anomalies) are dealt with in a timely manner and opportunities (positive anomalies) can be seized upon.

There are numerous algorithms and approaches to anomaly detection, however the techniques are not always applicable in the context of marketing analytics data because of its inherent seasonal and trend components. Anomalies are contextual in nature and the techniques used to develop the detection in one domain can rarely be used ‘as is’ in another domain, which makes this approach unique.

Signals that can be detected by the algorithm

These types of anomalies are well suited to the algorithm and will be raised as anomalous. As some metrics increasing in value should be considered a negative anomaly, we apply this logic for you so that you can set the notification policy that is right for you.

Grow to early in seasonal metrics

Some unusual noise

More noise than usual

Break down

Sudden growth

Elevated sustained growth

Pick

Activity when usually none

Signals that can't be detected by the algorithm

Every algorithm needs to make certain compromises in the approach and therefore has degraded performance across every type of data set. In complete transparency, the following list of scenarios are not well suited to this algorithm and it's important to be aware that not every single anomaly will be picked up.

Flat Signal

No Noise

Exponential

Negative Seasonal Anomaly

Negative Seasonal Anomaly

Anomaly Detection Algorithm Whitepaper

If you are interested in the research and statistical methods behind the algorithm you can access it below.

Anomaly Detection Whitepaper
See all categories

Anomaly Detection

Early detection of anomalies plays a key role in ensuring that issues (negative anomalies) are dealt with in a timely manner and opportunities (positive anomalies) can be seized upon.

There are numerous algorithms and approaches to anomaly detection, however the techniques are not always applicable in the context of marketing analytics data because of its inherent seasonal and trend components. Anomalies are contextual in nature and the techniques used to develop the detection in one domain can rarely be used ‘as is’ in another domain, which makes this approach unique.

Signals that can be detected by the algorithm

These types of anomalies are well suited to the algorithm and will be raised as anomalous. As some metrics increasing in value should be considered a negative anomaly, we apply this logic for you so that you can set the notification policy that is right for you.

Grow to early in seasonal metrics

Some unusual noise

More noise than usual

Break down

Sudden growth

Elevated sustained growth

Pick

Activity when usually none

Signals that can't be detected by the algorithm

Every algorithm needs to make certain compromises in the approach and therefore has degraded performance across every type of data set. In complete transparency, the following list of scenarios are not well suited to this algorithm and it's important to be aware that not every single anomaly will be picked up.

Flat Signal

No Noise

Exponential

Negative Seasonal Anomaly

Negative Seasonal Anomaly

Anomaly Detection Algorithm Whitepaper

If you are interested in the research and statistical methods behind the algorithm you can access it below.

Anomaly Detection Whitepaper