Level 4 · Which orders are unusually large?

A rule instead of a feeling

“Unusually large” needs a definition you can write down and repeat. The usual one is the 1.5 × IQR rule, the same rule a boxplot uses for its whiskers:

• compute Q1 and Q3, the 25th and 75th percentiles;
• the IQR is Q3 - Q1, the spread of the middle half;
• the upper fence is Q3 + 1.5 × IQR, the lower fence Q1 - 1.5 × IQR;
• values beyond a fence are outliers.

It works on skewed data like order values because quartiles, unlike the mean and the standard deviation, aren’t dragged around by the extremes they’re trying to detect.

In pandas

quantile accepts a list and returns one value per fraction, which unpacks neatly:

q1, q3 = order_value.quantile([0.25, 0.75])   # 87.94, 650.48
iqr = q3 - q1                                    # 562.53
upper = q3 + 1.5 * iqr                           # 1494.27
lower = q1 - 1.5 * iqr                           # negative: no low outliers possible
outliers = order_value[order_value > upper]
len(outliers)                                    # 370

Out of 4,295 completed 2025 orders, 370 are above $1,494.27, about 8.6%. That’s more than you’d see in a symmetric distribution, which is the long right tail again.

A seaborn boxplot of a single Series shows the same thing: sns.boxplot(x=order_value). Every dot to the right of the whisker is one of those 370 orders, and an axvline at the fence makes the rule visible.

Outlier is not the same as wrong

The next step is to look at who the outliers are, never to delete them on sight. Joining the segment shows that 256 of the 370 come from business accounts, which make up about a quarter of all orders. They’re unusual, and they’re some of the most valuable orders the store has. Dropping them would cut a large slice of revenue out of every total.

What you do depends on the question. For “what does a typical order look like”, report the median, which isn’t moved by them anyway. For revenue, keep them. For a model that would be thrown off by them, cap or segment them, and write down that decision and why. An outlier that is a real error, like a negative quantity, is a cleaning problem with its own rule.

Common mistakes

• Using quantile(25): quantile takes fractions, 0.25 and 0.75.
• Computing the fences on order lines instead of orders.
• Dropping outliers without checking what they are.
• Forgetting the lower fence exists. Here it’s negative, so it flags nothing, but say so.