Every Dataset Needs Two Averages, Not One

The average is one tidy little number that's supposed to speak for a whole group. Someone hands it to you and expects you to nod along, and most people do. That's exactly how the average keeps getting away with it.

Here’s a healthier reaction: squint at it like it just told you it “left the party early because it was tired.”

Say a nice person sells you a lottery ticket and assures you that “on average, you win $95 for every $100 you spend.” Wonderful. A cozy little $5 loss for a shot at a million dollars. Practically a public service, really. Where do you sign?

Except the average conveniently forgot to mention the fine print. Nobody actually walks away five dollars poorer. What really happens is that almost everyone loses their entire $100, and one lucky soul in many thousands strolls off with $1,000,000. Add it all up and sure, it “averages” to $95, technically, in the same way that you and Jeff Bezos are, on average, billionaires. True? Yes. Useful? Absolutely not.

This is the precise moment humanity stopped taking the average at its word and started asking the obvious follow-up: okay, but how spread out are the real numbers, and why do I feel like I’m being lied to?

Congratulations, you've just invented variance.

Variance is the number you bring in to interrogate the average. It measures how far the actual numbers wander off from where the average swears they all are. The method is almost insultingly simple: take each real number, subtract the average, and square the result so the negatives can’t sneak off and cancel everything out. Do that for every number, then average all those squared gaps. In other words, you take the average of exactly how much the average was wrong. Poetic.

So now you’ve got two averages on the payroll. The first one, the plain average, cheerfully tells you where the center of your data supposedly sits. The second one, variance, follows it around with a clipboard, quietly noting how much you should actually believe a word it says.