What can a rolling average show?

May 21, 2016 6:49 PM

Accounting filter: What can a rolling average reveal?

Industrial supply company. Our dashboard widget shows one year of a rolling monthly average for certain KPIs. Generally speaking, beyond mitigating seasonal spikes, what insight can a rolling average provide?

I haven’t provided specific KPI examples because I’m hoping to learn the essence of the rolling average.

posted by LonnieK to Work & Money (6 answers total)

You pretty much said it. A rolling or moving average is a low pass filter that removes spikes in the data allowing you to see the smoothed overall trend more easily.

A good example would be climate change temperature data. The temperatures from year to year vary widely and look quite noisy, but the overall trend is consistently up.
posted by JackFlash at 7:10 PM on May 21, 2016 [2 favorites]

Here is a good example from global temperature data. The black line shows that the average temperature for each year can vary up and down dramatically. But if you use the 5-year running average shown by the red line, the long term trend is much more obvious. The high frequency noise has been removed.
posted by JackFlash at 7:21 PM on May 21, 2016

I use control charts for my key measures. I don't use rolling averages because they can hide variation that may be significant in some way. A control chart helps you figure out whether the variation (or "wiggles") in your data are "common cause" variation characteristic of your processes or "assignable cause" variation that indicates something is up that might need attention. If I need to extract a summary of the "trend" i prefer to fit a curve rather than smooth the data with a moving average.
posted by cross_impact at 9:28 PM on May 21, 2016 [2 favorites]

Sometimes you can't use a control chart--when you expect the value to change. Ex., you would use a rolling average while you're trying to lose weight to smooth out the noise, then a control chart when you reach your goal weight to help maintain. Body weight has a lot of noise to it so you wouldn't want a graph without some kind of smoothing, ex. if you drank a big glass of water you would instantly have a big spike, but it usually doesn't fit to a nice curve at all.
posted by anaelith at 5:48 AM on May 22, 2016

Control charts are great for identifying anomalies in processes you expect to be relatively consistent. The classic use case is in manufacturing: You have a machine that makes widgets. On average, X widgets per day come out defective. One day 2X widgets are defective. Is that just bad luck, or is something wrong with the machine? Control charts are great for answering that question. They're very good for identifying significant changes (Is the increase in defective widgets statistically significant) but don't particularly help you figure out the magnitude of the change (how many defective widgets should we expect going forward?). They're also less good for things where you might expect an ongoing trend.

If you're just worried about seasonal variation, you usually see businesses do year-over-year comparisons. For example, in retail in Western countries the 4th quarter has all the holiday shopping, so comparing Q3 with Q4 of a given year isn't very meaningful, instead you generally compare Q4 of one year to Q4 of the previous year.

Rolling averages are good for finding trends in data which is noisy (it moves around more or less randomly) and granular (there's lots of data, so each individual data point isn't terribly important).
posted by firechicago at 6:38 AM on May 22, 2016 [1 favorite]

One important feature of a rolling average is that it's always out of date. As well as smoothing data, calculating a rolling average introduces lag.

For example, if you plot a monthly rolling average over daily data, you'll notice that any trend that's significant enough to show up in the rolling average will do so about two weeks after it shows up in the raw readings; significant peaks and valleys are similarly delayed, as well as being attenuated.
posted by flabdablet at 6:58 AM on May 22, 2016 [1 favorite]