The One Thing Worse Than Holiday Shopping? Monitoring Holiday Shopping

Everyone loves the holiday season, even the most cynical among us who feel it’s been distorted by a nefarious commercial plot concocted by greeting card companies. That being said, holidays are a nightmare for people responsible for business monitoring. As more holidays turn into mega shopping events, we see that shopping events turn into holidays – whether it’s Black Friday, Cyber Monday or China’s Singles’ Day.

Either way, holidays are notoriously hard for the professionals who are in charge of business monitoring. Whether you’re an e-Commerce site, a payments company or a restaurant (keep that example for later), making sure your business is running smoothly is based on understanding your ‘normal’. How much do you sell on a Sunday? How many transactions do you handle in the early morning in France? You rely on this ‘normal’ to understand if and when something is going wrong, regardless of the method you are using to detect it. 

The problem with holidays is that, on one hand, they are not normal – your business metrics behave differently during the holidays. Now let’s assume you have an alerting system. It monitors your ‘regular’ business behavior. Then comes a holiday. And all your alerts go crazy. One option would be to just turn off your alerting system so you don’t get bombarded with false negatives. Does that sound like a good idea? 

What you really want is to be able to compare the ‘not normal’ of this year’s holiday vs. the ‘not normal’ of last year. Because yes, you’re selling a lot more on Black Friday. But are you selling a lot more compared to last year? 

A common misconception is that business on the holidays is “the same, only more so”. That is not always true. Sure, retailers sell a lot more on Black Friday than a regular Friday. But it’s not only a difference in overall volume, it’s also a difference in behavioral patterns. Take our aforementioned restaurant. Assume it’s located in Istanbul. Each year during Ramadan, its Muslim patrons won’t coming in during the day, but at sundown orders come flooding in. This causes an increase in food consumption, only at very different hours than the normal. So it’s not just about volume, it’s also about changing patterns.

Last but not least, holidays are hard to monitor since they do not necessarily occur at a regular cadence.  You can’t just set alerts to suppress warnings every 365 days. Some are easy to predict because they occur on a specific date. However, consider holidays which are based on lunar calendars or holidays that are defined as happening on the fourth Thursday of November, like the American holiday of Thanksgiving. In the week or so leading up to that holiday, turkey sales accelerate exponentially until they’re sold out. But how do we prepare for these kinds of predicted patterns using what we already know.  

This is where we see more machine learning algorithms evolving to answer the challenge. While anomaly detection algorithms are becoming easier to use and access (running them at scale is a different story), some of the advancement in this field is based on the ability to learn an event’s behavioral patterns and then apply that learning to the next occurrence of the event. In order for an anomaly detection algorithm to properly handle such a unique event, we need to first train it to know how it should behave. Based on that training, the algorithm should be able to apply that learning as the new baseline which should be applied only when the next occurrence of the event happens. So, if I use the data from last Thanksgiving as my training set, the next Thanksgiving I will be able to use that as my baseline and only alert on the delta between the holiday occurrences and not the delta between the holiday and the day to day. So there is hope.

As we get ready to wrap up this year’s holiday season, we hope you manage to connect with loved ones,  and, as you go about holiday business, that you keep in mind those colleagues hard at work monitoring your business during the holidays.




About the Author:
Yariv Zur, VP Product at Anodot and oversees the company’s ML-based Autonomous Business Monitoring platform.