Rolling mean

What it measures

Nothing new, and that is the point: a rolling mean is a presentation choice, not a metric. Weekly efficiency numbers bounce hard (one week is roughly 60 plays), so a line chart of raw weekly EPA per play looks like a seismograph. Smoothing each point with the average of its recent window turns the seismograph into a trend you can actually read: is this offense getting better, worse, or holding steady?

How it is computed here

A trailing window: the smoothed value at week N is the plain average of the raw values from the last window weeks up to and including N. Early weeks warm up with whatever history exists (week 2 of a 3 week window averages weeks 1 and 2), so the line starts at the real first value instead of losing its opening weeks. The builder's rolling average control applies this to any line chart, with windows from 2 to 17 weeks.

How fast it stabilizes

This card is the stabilization story for everything else. The window size is a dial between responsiveness and reliability:

• Small windows (2 to 3 weeks) track real changes quickly (injury, new starter, scheme shift) but still carry plenty of noise.
• Mid windows (4 to 6 weeks) are the usual sweet spot for weekly team efficiency.
• Large windows (8+) are very stable and very slow; a real mid-season change takes most of the window to show up.

There is no correct setting, only an honest one for the question being asked.

How it gets misused

• Reading smoothed lines as raw data. Every plotted point mixes several weeks, so the "week 9 collapse" on a 5 week line may have happened in week 7. Our charts label the window for exactly this reason.
• Shopping windows until the story looks right. Trying settings until a trend appears is a way of manufacturing narratives; pick the window for the question, then keep it.
• Smoothing tiny histories. A 5 week average over a 4 week season is just the season average wearing a costume.