Let’s say you mistakenly change the price of a pair of jean shorts for a sale from $50.00 to $5.00. Many sales will go through, and at first, it may seem like the product is selling well because of the number of units being sold. However, when looking at the net revenue, it will tell another story. How do you ensure that a problem like this is identified and resolved swiftly? Anomaly detection systems act as a safeguard against problems like this by ensuring that such issues are detected as soon as possible.
Anomaly Detection is just as the name suggests; it is a process that mines through extensive data to trigger alerts if something is skewed or amiss. In the case of the jean shorts, an anomaly detection system would analyze the data, notice a discrepancy in net revenue, and then create an alert. However, the more important questions are: how does an anomaly detection system determine if a certain data point is an anomaly? How does it determine if something is normal? Although these questions may seem fundamental, there are a lot of things to consider when determining if something is an anomaly.
How Anomaly Detection Works
To begin working with an anomaly detection system, you first set certain upper and lower thresholds for each metric such as a return rate, units shipped, or recency of data points. From there, the system must identify anomalies comparing the current metrics with the historical data of the product. There are several factors to consider when determining how the anomaly detection system should run:
- Timeliness: How quickly does a company need an answer if something is an anomaly? If answers are needed slower or less frequently, the system will run through more data points and be more accurate. If answers are needed more frequently, they will be less accurate.
- Scale: The size of the datasets alter the way in which an anomaly detection system analyzes the data. Understanding the scale of the data is important for a software architect when making programmatic adjustments to the anomaly detection system.
- Rate of Change: If the datasets are constantly changing, then the system will need to utilize machine learning/adaptive algorithms to take certain changes into account. If the rate of change is slow, the system can collect data and determine what is normal from that dataset.
- Conciseness: How in depth do the metrics need to be measured? Do you just need an alert if an anomaly is detected or do you want reasons as to why it was triggered? And then, how in depth do you want the reasons to be?
Retailers incorporate anomaly detection software to help them notice issues in real-time and to learn from trends that otherwise would not be apparent when reviewing data on a smaller scale. As businesses grow, incidents (like the case of the $5 jean shorts mentioned earlier) go undetected unless there is an anomaly detection system to sift through the immense volume of data. Each incident is an opportunity to save money and create potential new business opportunities. Returns reduction, in particular, is one area of opportunity.
Anomaly detection is extremely important in assisting retailers in reducing returns, which is why we built an anomaly detection engine (ADE) within Chief Returns OfficerⓇ. By analysing metrics such as return rate, units shipped, or recency of data points, we are able to raise alerts on particular styles or product categories that are causing return issues. Retailers are then able to take actions on those alerts, such as fixing the sale price or adjusting the website copy or images based on customer feedback. But this is just one application of anomaly detection, which is a must-have and a strategic enhancement to your business.
This piece by Navjit Bhasin, Founder/CEO of Newmine, originally appeared on the Newmine blog.