Types of Anomalies
This article discusses the different types of anomalies that Auto Insights detects. Auto Insights defines an anomaly as “a value that sits outside the expected range.” But what does that actually mean?
As you upload data, Auto Insights uses a number of different algorithms (like ARIMA, S-ARMIA and others) to scan your data to find patterns and relationships between segments and measures and starts to learn about your data.
Auto Insights needs at least seven months of data to find anomalies.
It also identifies the historic pattern of your data. For example, your dataset includes a field named Revenue. Auto Insights will identify if Revenue has increased or decreased over time. It uses this information for two purposes:
1. To provide an overview of Revenue’s performance in Discover
2. To identify if the change between this period and the last period is in line with past performance, or if the change was unexpected.
Let's Explore Some Examples
1. Anomalies with more than 12 months of data
In this example, we see that revenue for Paper has decreased from Nov-Dec 2017 by -44% as well as in Nov-Dec 2018 by -11%, but has increased in Nov-Dec 2019 by 14%.
Auto Insights has identified that revenue for paper had experienced anomalous behavior between the latest period.
2. Anomalies with less than 12 months of data
In this example, we can see that the dataset only has 9 months of data. So how has Auto Insights identified that this value is "outside the expected range" if it can't compare it to the same period last year?
In this case, Auto Insights calculates the normal distribution of the data. Auto Insights will consider a value to be an anomaly if it is more than 1.5 standard deviations from the mean of the normal distribution.
3. Anomaly where Auto Insights says a measure "remained flat"
In this example, we can see Auto Insights tells us that Revenue remained flat between November and December 2019. Why is this an anomaly?
Remember that Auto Insights looks at the historic pattern of your data. You can see in this example, Revenue has historically decreased by around 39% for this product between November and December. However, between November and December 2019, Auto Insights observed that Revenue remained steady.
Therefore, as Revenue did not decline, Auto Insights has identified this as an anomaly.