Level 6 · Boss: the first-look report

Headline facts need precise definitions

“How many customers do we have?” sounds simple, but there are at least two answers in this dataset. len(customers) counts everyone who signed up. orders["customer_id"].nunique() counts everyone who actually placed an order. The difference, people who signed up and never bought, is a number marketing cares about a lot. When you report a headline figure, say which definition you used.

Step 1: distinct versus total

len(orders)                         # 12000 orders
orders["customer_id"].nunique()     # distinct customers among them

A customer with five orders appears five times in orders, so len overcounts customers. nunique() counts each ID once. This is one of the most frequent mistakes in junior reports, and one of the most frequent interview questions: “count customers, not orders”.

Step 2: shares instead of counts

Boards and managers want proportions, because they compare across time and across businesses of different sizes:

orders["channel"].value_counts(normalize=True).round(3)

normalize=True divides each count by the total. Round at the end: .round(3) keeps shares like 0.493, which is 49.3%. If you need percentages for a slide, multiply by 100 or format with f"{share:.1%}".

Before reporting shares, check that the column has no missing values (orders["channel"].isna().sum()). value_counts drops missing values by default, so the shares would quietly add up to 1 over only the known rows.

Step 3: a price summary

products["unit_price"].describe().round(2)

The result is a Series indexed by statistic name, so you can pull out single numbers:

summary = products["unit_price"].describe()
summary["50%"]     # the median price
summary["max"]     # the most expensive product

When you present this, lead with the median and the range, and mention that the mean is pulled up by expensive furniture. A summary that only says “average price $270” invites the wrong picture of a store where half the products cost under $154.

The habit behind this level

Whenever you get a new dataset, run this kind of profile before answering any question:

1. Row counts and the grain of each table.
2. Distinct counts of the key IDs.
3. value_counts(dropna=False) on every category column.
4. describe() on every numeric column.
5. isna().sum() everywhere.

It takes five minutes and catches most data problems before they reach a report: the duplicate customers, missing costs and odd quantities you’ll deal with in the cleaning chapter all show up in exactly these checks.

Common mistakes

• Using len(orders) or orders["customer_id"].count() for distinct customers. count() counts non-missing values, not distinct ones.
• Forgetting normalize=True and returning counts.
• Rounding the column before describe(), which changes the statistics. Round the summary instead.