Rolling Metrics

Raw metrics can be terribly noisy. Mornings are busier than evenings. Holidays create big dips. A sale creates an artificial spike.

Rolling averages smooth out all that noise by averaging the numbers over a moving time window, so you can see the actual trend. The window can be anything: days, weeks, months, etc.

It's not just the value for a single day. It's the average of the values in the window up to that day. So a 3-day rolling average for January 3rd is the average of January 1st, 2nd, and 3rd.

The rolling() method computes a function over a sliding window of N observations. To make each observation one calendar day, include missing dates first:

events["date"] = events["timestamp"].dt.floor("D")
daily_users = events.groupby("date")["user_id"].nunique().asfreq("D", fill_value=0)

# 7-day rolling average
rolling_7d = daily_users.rolling(7).mean()

Which prints something like:

7-day rolling average:
date
2024-01-01          NaN
2024-01-02          NaN
2024-01-03          NaN
2024-01-04          NaN
2024-01-05          NaN
2024-01-06          NaN
2024-01-07    13.000000
2024-01-08    14.142857
2024-01-09    14.857143
...
Name: user_id, dtype: float64

The first 6 values are NaN because there aren't enough prior days to fill the 7-day window! The min_periods parameter can relax that restriction and start calculating the average as soon as there's at least 1 day of data:

rolling_7d = daily_users.rolling(7, min_periods=1).mean()

7-day rolling average:
date
2024-01-01    10.000000
2024-01-02    11.000000
2024-01-03    11.000000
2024-01-04    11.750000
2024-01-05    12.400000
2024-01-06    12.500000
2024-01-07    13.000000
2024-01-08    14.142857
2024-01-09    14.857143
...
Name: user_id, dtype: float64