Follow Datanami:
October 18, 2019

Four Key Attributes of Advanced Anomaly Detection

Amit Levi


AI anomaly detection is just one tool in an arsenal that Gartner now refers to as augmented analytics, but it’s among the most important. By providing real-time intelligence into both positive and negative anomalies, AI anomaly detection offers opportunities to recapture revenue, capitalize on moment-to-moment trends, and sidestep brand damage.

But before we get into the four attributes of advanced anomaly detection, a couple of counter examples are in order. Let’s look at what happens when AI anomaly detection is not in place. At the macro level, you won’t see things like outages, bugs, or price glitches in your platform. This could lead to a situation where your customers know about an outage before you do. This isn’t a good look for your long-suffering support team. Your customers will become angry and churn, and your reputation will suffer.

Here’s what this looks like in real life:

NYSE Price Glitch Sows Widespread Confusion

On Monday, August 12, the New York Stock Exchange suffered a technical glitch that prevented traders from knowing the exact prices of the stocks they were trading. The stock exchange calculates national and global prices using three separate tapes, which are all combined to make a single stream of real-time information.

(Vintage Tone/Shutterstock)

Per the NYSE, this glitch was due to a network component failure. Failures in network components are often telegraphed by anomalies prior to a complete outage. These might include bad packets, increasing latency, or error messages. An automated anomaly detection solution could have flagged these anomalies beforehand.

At the same time as the NYSE experienced its network component failure, the investment company Vanguard lost its ability to update data on its own stocks. This led to an error in which its investment portfolio appeared to lose up to half its value. This error was directly related to the NYSE outage.

An advanced anomaly detection system has the ability to correlate related anomalies – which makes it that much easier to find the source of an outage. Given the data available, an analyst at the NYSE could have received an alert that looked like this (albeit in a much more technical visualization): “We’re currently experiencing an increase in bad packets and in latency, and we’re no longer able to communicate with Vanguard.” Using these data points, it would be that much easier to triangulate the source of an error.

Etsy Bug Provides Random Admin Access

In the meantime, a bug on July 8th accidentally gave random visitors administrative access to Etsy seller profiles. As part of a flawed A/B test, Etsy visitors were given the power to change some content on Etsy seller pages, alter the user interface, and gain detailed information on sellers. Since many Etsy sellers understandably don’t want strangers learning their personal info, a large fraction of them went into “vacation mode,” rendering them unable to complete sales.


Once again, there are a lot of potential metrics that would have allowed the Etsy team to quickly detect and resolve given an AI analyst insight into this issue. For example, it could have detected an abnormal spike in sellers closing up their shops. It could see that visitors might be spending an abnormal amount of time on seller pages. It might even detect a spike in the number of users with admin privileges (a useful metric to monitor in case of a cyberattack).

In any event, this bug frightened the Etsy user base and prevented them from making sales.

How Can AI Anomaly Detection Do What Others Can’t?

As the title of this article suggests, there are four ways in which anomaly detection gives you greater control of your business through proactive monitoring:

1. Real-Time Analysis

The longer an anomaly goes on, the more your customers and users are affected. You need to be notified as soon as an anomaly starts. You need to have all the information needed to investigate and resolve it as soon as you are notified about the anomaly. You can’t afford to use  batch algorithms in order to detect anomalies and get root cause information on it. Your anomaly detection system should give you both the anomaly and its potential correlations that let you remediate anomalies almost as soon as you find them.

2. Correlation of Anomalies

Your business runs hundreds of different applications. Getting data from those applications into a traditional analytics tool can be a frustrating process. What’s more, it becomes impossible to see whether an anomaly in one application is reflected in a different — but related — system.

Automated anomaly detection systems can  correlate anomalies between related metrics. Per the example from the NYSE glitch, it would have allowed an anomaly detection system to draw lines between an increase in latency, a decrease in transactions, and a communications outage with a corporate partner in order to help data scientists pinpoint the source of an error.

3. Works at Scale

One area where traditional tools and methods falter is that there’s only so many metrics they can measure. Not so for AI, which can simultaneously monitor thousands of KPIs. This is valuable due to the chance that metrics you wouldn’t ordinarily keep an eye on (see: number of users with admin privileges) might signal alarming anomalies.

4. Proactive Monitoring

The norm for your business may change day by day. As an easy example, a swimsuit store will probably see more orders during the beginning of the summer, and fewer as summer changes into fall. An anomaly detection system sees these changes and accounts for them as a normal pattern. For example, system models should be able to determine whether a sales slump at the end of August is a normal change in the baseline or a problem that quickly needs fixing.

Where Traditional BI Tools Fall Short

If you’re interested in finding anomalies and outliers in your ecommerce platform, you may be wondering why the usual run of BI tools tend not to stack up. In general, these tools can detect anomalies, but not really. Users need to set up manual thresholds for alerting – but as we’ve said, these thresholds change all the time in ways that don’t necessarily indicate that something is wrong.

Even when an anomaly does trip a threshold, it can take a long time for it to become apparent – long enough that your customers notice the impact. In other cases, a genuine anomaly will occur that doesn’t trip your manual thresholds. You’ll only discover the lost revenue opportunities after the fact.

Lastly, most BI tools only give you an average of all your metrics – which lets you ignore the anomalies hidden in the details. By analyzing thousands of metrics in real-time, AI analytics let you trace down problems that you’d otherwise miss. It’s a true industrial revolution in terms of maximizing revenue and customer satisfaction at scale.

About the author: Amit Levi is VP of product and marketing at Anodot. He is passionate about turning data into insights. Over the past 15 years, he’s been proud to accompany the development of the analytics market. Having held managerial positions in several leading startups, Amit brings vast experience in planning, developing, and shipping large scale data and analytics products to top mobile and web companies. An expert in product and data, his mantra is “Good judgment comes from experience and experience comes from bad judgment.”

Related Items:

Why Machine Learning Is Our Last Hope for Cybersecurity

Anodot Gains Patents for Anomaly Detection