Level 8 · Boss: the holiday season review

One table, four views of time

Every step starts from lines: one row per order line of a completed order, with its line_revenue and the order’s order_date. From there, time can be cut in two ways. Resampling puts rows into consecutive periods on a timeline (2 December 2025, the week ending on it, November 2025). Date parts pull one piece out of each date (its weekday, its month of the year, its quarter) so rows from different years can be grouped together. The boss needs both.

Date parts with .dt

The .dt accessor extracts parts of a datetime column:

d = lines["order_date"]
d.dt.year           # 2025
d.dt.quarter        # 1 to 4
d.dt.month          # 1 to 12
d.dt.month_name()   # 'November'
d.dt.day_name()     # 'Friday'

year, quarter and month are attributes (no brackets) and return numbers; month_name() and day_name() are methods and return text. For any other label, strftime formats a date with codes: "%Y-%m" gives '2025-11', "%b" gives 'Nov', "%B" the full name, "%a" a short weekday. On a DatetimeIndex the same methods are available directly, without .dt: table.index.strftime("%b").

Months in calendar order

Month names are text, and text sorts alphabetically: April, August, December… To keep calendar order, carry the month number next to the name and sort by the number:

parts = lines.assign(month_num=lines["order_date"].dt.month, month=lines["order_date"].dt.month_name())
parts.groupby(["month_num", "month"], as_index=False)["line_revenue"].sum().sort_values("month_num")

Grouping by both keys costs nothing (each number has one name) and gives you a column to sort on. Another option is an ordered categorical, pd.Categorical(names, categories=[...], ordered=True), but a month number is simpler and survives every merge.

Averaging a month over three years needs two steps: first sum per year and month, then average those totals. Averaging lines directly would give the average order line, not the average month.

Rolling windows and growth, history first

The best week of 2025 is a rolling 7-day sum over a daily series, built on the full history and filtered to 2025 afterwards. It ends on Thursday 4 December 2025 at $104,267.39, when the window holds the whole Black Friday to Cyber Monday run. In 2024 the best window was bigger, $117,208.44.

The same rule applies to year-over-year growth: pct_change(12) needs the 2024 months in the series, so select Q4 2025 only at the end. The answer is uncomfortable: November 2025 revenue fell 8.4% on November 2024 although order counts grew 13.7%, the kind of result you report plainly, with the order count next to it.

Shares of each year

transform("sum") returns the group total on every row of the group, so a quarter’s share of its year is one division. Q4 brought 0.375 of 2023 revenue, 0.381 of 2024 and 0.328 of 2025.

Common mistakes

• Sorting month names alphabetically instead of by month number.
• Writing dt.month_name without brackets, which gives a method, not names.
• Filtering to 2025 before rolling or pct_change(12), which turns the start of the period into NaN.
• Dividing by the grand total of all years instead of each year’s own total.