Level 3 · A 7-day average for the dashboard

Why smooth a daily series

Daily numbers are noisy. Harbor & Pine’s orders swing with the day of the week, with paydays and with promotions, so a daily chart looks like a seismograph. A moving average (also called a rolling average) replaces each day with the mean of a window of days ending on it. A 7-day window covers every weekday exactly once, so the weekly rhythm cancels out and the trend shows.

rolling(n)

rolling(7) slides a window of 7 rows along the series; the aggregation after it runs on every window:

daily = orders.set_index("order_date").resample("D").size()
daily.head(8).tolist()
# [1, 8, 6, 4, 10, 9, 5, 8]
daily.rolling(7).mean().round(2).head(8).tolist()
# [nan, nan, nan, nan, nan, nan, 6.14, 7.14]

The 7th value, 6.14, is the mean of the first seven days (43 orders / 7). The first six values are NaN because a full window of 7 rows doesn’t exist yet. If you’d rather get a value from the first day on, pass min_periods: rolling(7, min_periods=1) averages whatever rows are available, so the first three values become 1.0, 4.5 and 5.0. That’s convenient but misleading at the start of a chart, which is why the default is strict.

.sum(), .max(), .median() and .std() work after rolling too. A rolling 7-day sum is “orders in the last week”, another common dashboard number.

The window counts rows, not days. That’s why the daily series comes from resample("D"): it has exactly one row per calendar day, zeros included, so 7 rows really are 7 days. On a series with missing days you’d use a time-based window like rolling("7D") instead.

Window first, filter second

Jordan only wants November and December. If you filter first, the window has nothing to look back on:

daily.loc["2025-11-01":].rolling(7).mean().head(7)   # six NaN, then 17.71

Compute the rolling mean on the full history and cut the date range at the very end. Then 1 November 2025 averages 26 October to 1 November: 14.29.

What the holiday season looks like

On Black Friday, 28 November 2025, orders jump from 24 to 56. The daily count spikes for four days (56, 51, 58, 67) and drops back to 24 on 2 December, but the 7-day average keeps climbing for a few days, because the spike is still inside the window. It peaks at 44.14 on 4 December, the last day whose week still contains the whole Black Friday weekend and Cyber Monday. A moving average always lags: it tells you the level of the last week, not of today.

Common mistakes

• Filtering the date range before rolling, which leaves NaN at the start of the period.
• Rolling over raw orders instead of a daily series: the window is 7 orders, not 7 days.
• Using min_periods=1 without saying so, which makes the first days look calmer than they were.
• Reading the peak of a moving average as the peak day. It lags the real spike.