"The average customer spends $74"
It sounds like a fact. It is actually a decision. Lumen Skincare, the direct-to-consumer brand we follow through this whole course, has thousands of orders - and someone chose to squeeze them all into one number. Change the choice and the story changes: the typical order is closer to $63, a handful of $300 gift sets drag the average up, and "the average customer" is a person who may not exist. Descriptive statistics is the craft of summarizing data without lying with the summary. Get it right and you read every dashboard with clearer eyes.
The center: mean, median, and when the average lies 8 min live
There are three ways to name the "middle" of a set of numbers, and on skewed data they disagree - sometimes a lot. Knowing which one a report used, and which one it should have used, is the highest-leverage reading skill in this whole session.
LiveMean, median, mode - what each one actually means3 min▶
Three different answers to "what is the middle?", each right for a different job:
- Mean (the average) - add everything up, divide by the count. Great for totals and budgets (mean order x orders = revenue). Easily yanked around by a few extreme values.
- Median - line every value up and take the one in the middle. Half are below, half above. Barely moves when a few giants show up, so it is the honest "typical" on skewed data like income, order value, or time-on-site.
- Mode - the most common value. Useful for categories ("most orders are the $39 cleanser") more than for continuous money.
A Lumen deck says "average order value is $74, up from $71." A director wants to fund a campaign aimed at "the $74 customer." But the median order is $63 - most customers spend less than the average, and the lift came from a few big gift-set buyers over the holidays, not a broad shift. Targeting the mean would have chased a customer who is really a statistical mirage. Asking "mean or median?" reframed the whole plan.
LiveThe one question to ask about any average3 min▶
When someone quotes an average, ask: "is this data skewed, and if so should we be looking at the median?" Anything money-, time-, or count-shaped is usually skewed - a floor at zero and a long tail of big values. For those, the mean overstates the typical case.
- Symmetric data (heights, measurement error, many natural quantities) - mean and median nearly agree, use either.
- Right-skewed data (order value, salaries, session length, revenue per account) - median for "typical", mean only for "per-head totals".
- A single headline number hides the spread entirely. Part 2 is about getting the spread back.
Spread and shape: how much does it vary? 7 min live
Two teams both average 3.2% conversion. One is steady at 3.0-3.4% every week; the other swings from 1% to 6%. Same center, completely different reality - and a single average would have hidden it. Spread is the second half of every summary, and the boxplot is the fastest way to see it.
LiveRange, IQR, and standard deviation - three ways to say "how spread out"3 min▶
- Range - biggest minus smallest. Simple, but one freak value blows it up.
- Interquartile range (IQR) - the width of the middle 50% (Q3 minus Q1). Ignores the extremes, so it is the robust partner to the median. The box in the boxplot above.
- Standard deviation (SD) - the typical distance of a value from the mean. The partner to the mean, and the number that powers almost all the inference later in this course. A bigger SD means a wider, less predictable spread.
Two Lumen ad channels both return an average $74 order. Channel A's SD is $20; Channel B's is $70. Channel A is a dependable machine; Channel B is a lottery of tiny orders and occasional whales. Same average, opposite risk profile. A leader who only reads the mean funds them as if they were identical.
Self-studyShape: skew, and the percentiles you will actually be quoted4 min read▶
Shape is the third question. Most business data is not a neat bell - it is skewed, with a long tail on one side.
- Right-skewed (tail to the right) - order value, income, response time. Mean > median. The common case.
- Percentiles - "the 90th percentile order is $180" means 90% of orders are at or below $180. This is how you will hear tail behaviour described - especially for latency ("p95 page-load") and spend.
- Why it matters: service-level promises, capacity planning, and pricing all live in the tail, not at the average. "Average latency is fine" can coexist with "p99 is a disaster."
This week ◐ 25 min total
- Find one "average" in a recent report from your org. Ask: is this data skewed? Would the median tell a different story? Write one sentence on which you would trust.
- Hunt a hidden spread. Find a metric quoted as a single number (an average handle time, an average deal size) and ask your team for its range or its distribution. Notice how much the single number hid.
- Read one boxplot or histogram in a dashboard you already have. Name the center, the spread, and whether it is skewed - out loud.
- Optional: pick a metric and decide, in advance, whether mean or median is the right summary for it. You will defend that choice in Session 2.
Three questions before you go 🎯 ◐ 90 seconds
1 · Lumen's mean order value is $74 and the median is $63. What does that gap most likely tell you?
Mean > median is the signature of a right skew. A few large orders lift the average, so the median ($63) is the more honest "typical" order.
2 · Two channels both average a $74 order, but channel A has SD $20 and channel B has SD $70. The safest reading is...
The average hides the spread. A larger standard deviation means wider, less predictable outcomes - a real difference in risk that a single mean would conceal.
3 · A dashboard shows "average page-load is 1.2s" and everyone relaxes. What should a careful leader still ask?
Averages hide tails. A fine mean can sit next to a terrible p95/p99, where a meaningful group of users has a bad experience. Service quality lives in the tail, not the average.
What this session covers
This leader session teaches the reading skills behind descriptive statistics - the ~80% you need to interpret any summary honestly. The hands-on computation (numpy, pandas, matplotlib) lives in Builder Session 1; the sources below go deeper.