Level 9 · Boss: is it the channel or the customer?

Splitting by the second factor

The board compares channels. But a channel’s average basket depends on who shops there. If business customers preferred one channel, that channel would look better for a reason that has nothing to do with the channel itself. The check is to split by the second factor and see where the differences really are.

A pivot with channels down and segments across does exactly that, and a heatmap makes it readable:

grid = per_order.pivot_table(index="channel", columns="segment", values="value", aggfunc="median")
sns.heatmap(grid, annot=True, fmt=",.0f", cmap="Blues", cbar_kws={"label": "Median order (USD)"})

Use a sequential palette such as Blues for values that are all positive, light for low and dark for high. Keep the diverging coolwarm for numbers with a sign, like correlations.

Spread, not only the middle

A median per cell can hide very different spreads, so draw the boxes too. hue="segment" splits each channel’s box in two:

sns.boxplot(data=per_order, x="channel", y="value", hue="segment", order=order, showfliers=False)

showfliers=False hides the outlier dots so the boxes aren’t squashed by a few huge orders. Say so on the axis label, since you’re hiding data from the picture (not from the numbers).

A dashboard in seaborn

seaborn functions take an ax= argument, so a grid from plt.subplots holds seaborn charts as easily as matplotlib ones:

fig, axs = plt.subplots(1, 2, figsize=(13, 4.5))
sns.heatmap(grid, annot=True, fmt=",.0f", cmap="Blues", ax=axs[0])
sns.boxplot(data=per_order, x="channel", y="value", hue="segment", showfliers=False, ax=axs[1])
plt.tight_layout()

Exporting and writing it up

Colleagues get the table as a file: to_csv(path, index=False) writes CSV, to_excel an Excel sheet (it needs the openpyxl package), to_parquet a compact typed file for large data. Without a path, to_csv returns the text, which is handy for checking what you’d write.

Then the sentence. Medians by channel run from $249 on the marketplace to $263 on the web, a difference of about 5%. A business order’s median is about 3.8 times a consumer order’s. What: the channel hardly matters. How much: $249 to $263, against 3.8 times. So what: target by segment.

Common mistakes

• Comparing channels without splitting by segment, and crediting the channel for its customers.
• A diverging palette for all-positive values, which suggests a meaningful midpoint that isn’t there.
• Writing numbers into the takeaway by hand instead of formatting them from the data.