What is Mean Squared Error? — one analogy

1The analogy

Picture spending the whole day pouring water into measuring cups, trying to hit a marked line each time. Some cups end up a notch over the line, some a notch under, and every so often one is off by three whole notches. At the end of the day, the goal is a single number summing up how accurate the pouring was.

Simply adding up how many notches each cup missed by doesn't work. A cup one notch over and a cup one notch under cancel each other out. Both cups were off, but added together they look perfect.

So the notches get squared before adding. One notch off is a penalty of 1, two notches is 4, three notches is 9. Overshoot or undershoot, a penalty is left behind either way, and a badly missed cup carries a much heavier one. Add up all those penalties and divide by the number of cups, and that's Mean Squared Error.

2In detail

First, measure how far off each cup landed

There's a target value and there's the value actually poured. The distance between them is that cup's error. Whether the cup went over or under the line, the distance itself looks the same either way.

Adding these distances up as they are lets direction cancel direction out. If half the cups overshot and half undershot in a day, the total comes out near zero, which would say a shaky hand was somehow perfect all day. Something has to strip out direction and keep only size.

There are two ways to strip out direction: drop the sign and just count the distance, or square it. Mean Squared Error picks the second.

Why squaring, specifically

Squaring adds one more property while it's at it: a small miss shrinks even smaller, and a big miss grows much bigger. A cup one notch off carries a penalty of 1; a cup three notches off carries a penalty of 9 — three times the miss, nine times the penalty.

That's the attitude wanted in most cases. A cup one notch off is barely noticeable if you drink from it, but a cup three notches off is unusable. The pain of a mistake usually doesn't grow evenly with how far off it is — it grows much more steeply.

This property is useful during training too. Because the penalty rises steeply around a big miss, it's obvious how far and in which direction to correct the model. As the guess gets closer to the target, the penalty flattens out, so the finish is gentle too.

Averaging makes the count irrelevant

Just adding up penalties makes a day with more cups automatically score higher, and that rules out comparing today with yesterday. Dividing by the number of cups at the end fixes that. The value becomes "the average penalty per cup," and it can be compared whether there were ten cups or a thousand.

The three words in the name are the calculation steps in order: error is the distance, squared means squaring that distance, and mean is dividing by the count at the end. Knowing why the steps go in that order, rather than just memorizing them, makes other penalties easier to read too.

One cup can ruin the whole day

The biggest property of this penalty is how badly it's shaken by a stray value. Say ninety-nine out of a hundred cups landed within one notch, and one cup was off by ten. The combined penalty from those ninety-nine cups can still be smaller than the penalty from that single cup.

That property can be an advantage or a trap depending on the situation. If a big miss is genuinely a big disaster, penalizing it this way is the right call. But if the data has a value that was simply recorded wrong mixed in, the model gets dragged toward getting that one bad value right at the expense of everything else.

That's why the data gets looked at before this penalty is used. Confirm what a stray value actually is, clean it up if it's not something that really happened, and only then start training. If a stray value is real but its pull is too heavy to live with, a gentler penalty that counts distance plainly, without squaring, is sometimes used instead.

3More precisely

Mean Squared Error is hard to read on its own. Because the distance got squared, the units are squared too. In a problem about hitting a target in meters, the penalty's unit isn't meters. That's why results shown to people usually come with the square root taken, bringing the value back to the original unit. That number reads closer to "how many meters off, on average."

The difference from a penalty that counts distance plainly is worth knowing too. Counting distance plainly keeps a big miss from gaining extra weight, so it stays much calmer around a stray value. But the penalty then changes abruptly near the target, which makes training struggle to settle cleanly. Neither approach is simply correct — the choice is about which kind of mistake should hurt more.

The analogy has a limit too. A measuring cup shows its line, so a mistake can be corrected mid-pour. A model only receives its penalty after the whole guess is poured out. And this particular penalty is a ruler built for number-guessing problems. Applied directly to a problem about sorting things into categories, training doesn't go well, so a differently shaped penalty is used there instead.

4Try it yourself

TensorFlow Playground ailearn.space Watch the loss number in the top right shrink as training runs

5Common misconceptions

It's easy to think a low Mean Squared Error always means a good model, but actually removing a single badly-missed value can shrink the number a lot on its own.
It's easy to think the raw penalty value can be read directly as "off by this much on average", but actually the units are squared, so only the square-rooted value reads that way.
It's easy to think squaring is just there to get rid of negative numbers, but actually it's also there to make a bigger miss carry a heavier penalty.

7One-line summary

In shortMean Squared Error is a penalty that squares how far off each guess was and then averages them, so a single badly-missed value can drag the whole score around.

Did this analogy help?

Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02