Anomalies

Overview

Anomalies help marketers stay informed of statistical irregularities in the campaigns of connected ad platforms. hawke.ai monitors the fluctuations in key metrics of your connected digital platforms and when one of those metrics is statistically different from the expected range, the platform will trigger an anomaly for that campaign or platform.

Navigating Anomalies

The updated version of anomalies, released in April 2022, has been designed to mimic many of our other helpful features. At the top you’ll see the ability to toggle between anomalies that have been generated, as well as view those anomalies that have been dismissed by yourself, or by another user.

You will also see the total quantity of anomalies currently displayed with a further breakdown of their individual trends (positive or negative). We also give you the ability to adjust the date range of the anomalies displayed, giving you up to 12 weeks of helpful anomaly data.

Filtering Anomalies. We make it easy to filter your anomalies based on platform and/or by trend (positive or negative).

As you navigate down the page, you will see a collection of individual ad campaigns, as well as complete platforms:

Each of these parent campaigns or parent platforms contains:

• Icon of platform and name of campaign (or name of platform for complete platform anomalies)
• Date range of the values used for the metrics displayed
• Dismiss button (top right) to move the anomaly to the Dismissed tabCollection of metrics with values for above-mentioned date range, comparison to the previous week, and lightning bolts noting the metrics with active anomalies
• Anomaly cards for each of the identified anomalies (more information on anomaly cards below)

Anomaly Cards. To provide further context as to why the performance of a campaign is changing, each parent campaign or parent platform will contain at least one, and sometimes many, anomaly cards.

Each anomaly card contains:

• Name of the metric being displayed (specific to the parent)
• Note on how the recent trend compares to the expected range
• Plot showing up to 7 weeks of data with a thicker faded line noting the expected range – hovering on this line will show the exact values for each week
• Total for the specific metric for the previous week (matching the data range of the parent)
• Expected range for the specific metric

Defining Anomalies

hawke.ai’s anomaly algorithm monitors the health of performance metrics for each campaign of your connected ad platforms. This data is tracked weekly and compared to historical norms for each specific performance metric. When the difference between the value for the previous week differs significantly from the expected value, hawke.ai will identify this as an anomaly.

To determine the expected value, hawke.ai uses a neural network with unique datasets for each individual metric. This dataset includes the value for the past few days, the same day the previous month, the same day the previous year, as well as values associated for each day of the week.

Depending on the metric and severity of the difference (calculated based on standard deviation from the expected value), Morphio will color-code the anomaly accordingly.

Light green to dark green = positive performance change for the specific metric, with the shade of green changing based on the statistical difference.

Orange to red = negative performance change for the specific metric, with getting darker as the change becomes worse.

Dismissing (and Restoring) Anomalies

You have the ability to dismiss anomalies that you no longer want to track. Simply click on the “Dismiss” button in the top right corner of the campaign or platform, which will then move that card to a “Dismissed” tab on the same page. From there, dismissed anomalies can easily be reviewed and even restored back to the main section.

FAQ

Q: Some of the metrics show a large week-over-week changed, but have not triggered any anomaly. Why?
A: While week-over-week changes are a component of our anomaly algorithm, many other data points are used to determine the expected range for a metric which will then trigger an anomaly only when the previous weeks’ value is outside of that range.