With the online holiday shopping season heavily underway and expected to reach $83 billion this year, online stores need to run smoothly, including having prices in order. It may sound basic, but simple computer glitches can cause pricing errors that may excite consumers, but cause major problems down the line. For example, for a few hours this summer, visitors to Best Buy’s web site were able to purchase $200 gift cards for only $15 dollars. The company resolved the issue only after hundreds of buyers took advantage of the (erroneous) bargain and had to be reimbursed.
Similar apparent 'deals' happen very frequently. United Airlines sold free (50 GBP) first class tickets; Walmart sold $1000 treadmills for $33. These are perfect examples of delayed business insights that can cost companies millions of dollars.
So how do companies go about detecting and resolving such issues? One of the ways companies try to shorten the time needed to solve such cases is by using traditional Business Intelligence (BI) and monitoring tools. These include creating dashboards, reports or static thresholds. This approach poses a few problems:
Heavy maintenance
In the business world, millions of metrics may be collected. Humans typically will prioritize which metrics to track and which to ignore.
Considering the BestBuy example, the company most likely chose not to monitor the purchase behavior for that specific $200 gift card. Since they (and most other ecommerce businesses) need to monitor thousands of products, they cannot set alerting thresholds for every one of them.
Differentiating between Normal and Abnormal Metrics
Let's take a look at the metrics that you do choose to track: you need to create dashboard and reports on demand. However, these are created based on what you want to measure. What happens when you don't know or are missing a metric? Accurate alerts from these metrics are very difficult to achieve: it requires setting static thresholds. How can you determine what is normal behavior and what is not? Metric normal behavior can change dramatically over time – during the day, week etc. This factor results in numerous false positive alerts.
Business Insight Latency
Most significantly, traditional monitoring tools have inherent business insight latency: they will not show a real-time status – which means that you will discover business problems too late. Referring again to the BestBuy example, although the problem was indeed discovered, this occurred only a few hours later, and only after hundreds of people had already purchased the gift card for $200.
From Reactive to Proactive Monitoring
So how can you change a reactive business monitoring method to a proactive one? The answer is by using an automated anomaly detection system. This method will allow your business to track millions of metrics with minimal maintenance, get accurate alerts based on determining normal metric behavior vs. understanding of abnormal behavior, and gain real time insights when you really need them. With an automated anomaly detection system, you can actually identify issues as they happen, before they impact your business dramatically.
Anomaly Detection could have assisted BestBuy in the following ways:
1. Automatically learn normal behavior
The system will learn the normal behavior for “number of $200 gift card purchases” and “revenue from $200 gift card” metrics. The graph below shows a typical ecommerce retailer using Anodot’s automated anomaly detection. Notice that the behavior is seasonal. This means that it changes at different times of the day, on different days such as holidays or weekends. As you can see in shaded area in the graph, an automated machine learning system can cope with dynamic behavior.
2. Quickly Identify and alert on abnormal activity
The image below indicates how an anomaly detection-based system would discover the abnormal behavior much earlier than one based on static-thresholds.
3. Metric correlation to gain the right context
By correlating multiple metrics automatically, the automated anomaly detection system can send one alert with metrics that are more meaningful together. In this way, the operator would have seen immediately that there was unusual activity around the “number of $200 gift card purchases.” Without the –proper context, a steep increase in number of gift cards purchased would be perceived as a positive incident. However, a concurrent drop in the “revenue from $200 gift card” metric (see image below) would immediately indicate that something is going wrong; and would have dramatically reduced the time to resolution.
To summarize, using an automated approach, BestBuy and other ecommerce businesses can resolve such price glitches much faster. This scenario is just one of many in which an anomaly detection system would have been effective. Anomaly detection helps companies gain critical real-time business insights and optimize revenue.
Learn more about how you can achieve real time business insights through Anomaly Detection by visiting Anodot.com.
I recently spoke about automated anomaly detection in more depth at the CTOSummit, in San Francisco. Feel free to browse through the slides.
