iPhone 7, Headphone Jacks, and How Apple Hijacks Time-Series Data Trends

Posted by Alex Slotnick on Sep 7, 2016 5:40:04 PM

All ye faithful: the iPhone 7 is nigh, just as all tech blogs foretold. This afternoon, Apple kicked off September in the same way it has for the past several years, by making a product announcement and adding new technology to their market-warping and trend-setting lineup. As early Fall has become synonymous with Apple’s unique brand of press conference, these announcements have attained a reputation as faux-holiday, demanding live coverage across the internet, from Wired to The New York Times to The Guardian to The Times of India.

The iPhone 7, from Apple Press Info. For editorial purposes only.

Whether or not you’re one of the many, many viewers who tuned in for Apple’s press conference, a massive contingent of the Internet is always part of the audience. That means productivity at companies changes for a few hours, online traffic shifts, search terms become lopsided toward Apple, etc. And from a time series data standpoint, at VividCortex we find those phenomena very interesting. Particularly, we notice that as people’s online behavior changes, time-series algorithms such as the Holt-Winters double exponential smoothing method will more than likely be messed up for weeks to come, thanks to the massive dent put into traffic and productivity. We can expect for these algorithms to produce massive spikes and dips as they attempts to account for the big behavioral aberrations caused by today’s Apple conference. 

How does exponential smoothing respond to a spike like that?

Here's a visualization of how the annual bump in Google traffic around Apple's annual conference upsets the Holt-Winters model over a 5 year period:


This is all related to the time-series concept of seasonality: as time passes, it’s often safe to project and assume certain cyclical patterns. The earth’s seasons are an obvious (painfully obvious) example of seasonality in cultural and climate-related terms, but the concept also applies to any system for which you can predict certain behaviors or events happening at a distinct, predictable moment. 

Of course, one’s capacity for using something like the Holt-Winters method effectively is dependent on the length and completeness of a cycle of time within their model. If you want to be able to show that an anomaly happens once a year, at the same time each year, your model must have units of time big enough to contain that entire anomaly, plus the bigger picture within which it’s meant to fit. If you don’t have that, it’s impossible to show a cycle, it’s impossible to establish seasonality, and it’s therefore impossible to adapt for future projections.

People try to use algorithms like Holt-Winters for purposes such as anomaly detection and capacity planning, but when an event causes extreme activity of one kind or another (like, say, when an Apple conference occurs), it can really mess up those models. Catching such yearly events requires a very long season, but most people’s seasons aren’t quite that long.

Of course, in the world of the web and tech companies, there are far more season-warping events than just Apple conferences. Black Friday is another great example for retailers: once a year, for a day or two, all other retail activity models can be thrown out the window. So what can you do, if you know your model isn’t set up for such a hiccup? Naturally, we recommend some solid, deeply insightful monitoring tech — you can try a free trial of VividCortex to see what we mean — but preparation of any kind is the first step.  

Recent Posts

Posts by Topic

see all