Traditionally, anomaly detection systems focus on time-series anomalies. These are anomalies that are flagged in your metrics when there is a change in performance over time. An example of a time-series anomaly might be when:
- The click through rate of a campaign is 50% lower than it was the week before
- The average-order-value is significantly lower (or higher) one month compared to another
- One of your pages taking twice as long to load than normal
Time-series anomalies let you know that something has gone wrong during that period that needs further investigating. They are the generally the simplest anomalies to identify and track. However tracking anomalies by time-based changes only alert us to anomalies that show up over time, without giving any insight into the anomalies that appear between different dimensions, known as comparative anomalies.
What are comparative anomalies
Comparative anomalies aren’t anomalies that show up over time, they are anomalies that appear between different dimensions, showing discrepancies between the performance of specific metrics. For example, an anomaly that appears between dimensions might be when:
- One of your campaigns performs consistently poorly in one specific city, while doing well in other cities
- A product always has at least 70% fewer orders than average
- A page has always loaded slowly
When you’re only using a time-based anomaly detection system, anomalies which may have consistently under-performed won’t be flagged because it’s not unusual behaviour when looked at over time. But the fact that it’s constantly under-performing is vital information that will help improve digital performance.
Identifying comparative anomalies
To identify these comparative anomalies yourself, you have to scan through lists of hundreds of metrics across multiple dimensions, and use your eyes conduct analysis. This will not only be time-consuming, but it’s also likely that you will make mistakes when looking through that many numbers. Additionally, statistically insignificant data – such as a free product that was given out as part of a promotion or a product with just one sale – will make it even more difficult to understand the anomaly’s significance.
For example, take a look at of one of our e-commerce client’s* product list performance report on Google Analytics, a report which has over 3,000 products listed. Identifying comparative anomalies in this list week-in-week-out would take hours upon hours.
But if you can’t find these anomalies, some metrics will constantly under-perform, damaging your bottom line. That’s why we’ve created a comparative anomalies feature on the Millimetric platform, to help teams stay on top of anomalies that appear between dimensions.
Identifying comparative anomalies with AI
For example, here’s a comparative anomalies graph created on the Millimetric platform that plots an e-commerce client’s* data across Product Detail Views and Buy-to-Detail rate, identifying the average KPI as well as highlighting the ones that fall above and below ‘normal’ range.
Using this graph, the client can immediately see which products aren’t selling well, and the relationship this has to product views. This is information they can then use to decide which products to keep and stop selling, what to include in advertising alongside many other digital performance decisions.
The comparative anomalies feature is now available to Growth and Enterprise Millimetric users. Sign up for a Growth plan, or a 7-day free trial, on the Millimetric platform or contact us on firstname.lastname@example.org to find out more about the Enterprise plan.