Level 8 · How the channel mix shifted

What a total is made of

“How much of the total comes from each part?” is a composition question, and the chart for it is a stacked bar: one bar per period, split into segments. When the totals differ a lot between periods but you care about the mix, stack shares, so every bar is 100% tall and only the segments move.

Counts, then shares

pd.crosstab counts rows for each pair of values:

year = orders["order_date"].dt.year.rename("year")
counts = pd.crosstab(year, orders["channel"])
#        app  marketplace   web
# 2023   819          464  1888
# 2024  1366          601  2079
# 2025  2135          702  1946

To get shares within a year, divide each row by that row’s total. sum(axis=1) adds across the columns, giving one total per year, and div(..., axis=0) divides row by row:

shares = counts.div(counts.sum(axis=1), axis=0).round(3)
#         app  marketplace    web
# 2023  0.258        0.146  0.595
# 2025  0.446        0.147  0.407

Stacked and grouped bars in pandas

ax = shares.plot(kind="bar", stacked=True, rot=0)   # stacked: composition
ax = counts.plot(kind="bar", rot=0)                  # grouped: compare channels side by side

Grouped bars answer “which is bigger?” within each year. Stacked bars answer “how big is the whole, and what is it made of?”. rot=0 keeps the year labels upright.

In plain matplotlib a stack is built by hand: each segment starts where the previous one ended, which is what bottom= does:

fig, ax = plt.subplots()
bottom = np.zeros(len(shares))
for channel in shares.columns:
    ax.bar(shares.index.astype(str), shares[channel], bottom=bottom, label=channel)
    bottom += shares[channel].values
ax.legend()

Reading the shift

In 2023 the website took 59.5% of orders and the app 25.8%. By 2025 the app had 44.6% and the website 40.7%: the app passed the website in order count. The marketplace barely moved, 14.6% to 14.7%. Only the bottom segment of a stack can be compared exactly, because the others float, so put the segment your story is about at the bottom. With only three parts of one total, a pie per year would also work, but three pies side by side are harder to compare than three bars.

Common mistakes

• Dividing by column totals (counts / counts.sum()) instead of row totals.
• Stacking raw counts when the question is about the mix: the growing totals hide the shift.
• Too many segments. Past five or six, group the small ones into “other”.
• Forgetting the legend, or leaving it on top of the bars. bbox_to_anchor=(1, 1) moves it to the right of the chart, where it covers nothing.