Forecasting and Baselines
By the end of this lesson, you should be able to: explain why every "it went down" needs a baseline, build simple baselines that handle the weekly cycle, score them, and use a prediction range to decide whether a change is real.
Every comparison is a forecast
When you say "DAU fell 5%", you're comparing today with something: yesterday, last week, the same day last year. That something is your guess of what today should have been. It's a forecast, whether or not you call it one.
That guess is called a baseline. A good baseline makes most "why did it drop?" questions disappear, because often the answer is "it didn't, it's Saturday". A bad baseline creates alarms that aren't real and hides drops that are.
So before you look for why something moved, work out what it should have been.
The week has a shape
Most products have a weekly cycle, and it's often large. Here's English Wikipedia's daily views from September to November 2025, averaged by day of the week, split by desktop and mobile web.
Desktop views peak on weekdays, when people are at work and school. Tuesday averages 27% more than Saturday. Mobile does the opposite and peaks on Sunday.
Neither pattern has anything to do with Wikipedia's product. It's just the calendar. If you compare desktop views on a Monday with the day before, you'll see a jump of around 20% almost every week.
Notice too that the two patterns partly cancel out. Add desktop and mobile together, and the weekly swing is much smaller. That's the mix effect from lesson 13: a blended number can hide a strong pattern in each part, so build baselines on the parts you actually care about.