Because it gives more weight to one big error then to multiple small ones with the same sum.
We want the errors to be noise and not systematic. Noise usually has a gaussian distribution. And in a gaussian distribution multiple small values are more likely than one big one.
So Predic1 was better then Predic2? No. Because correctly predicting the one outlier shows more predictive power then staying close to the average. Therefore we use SumOfSquerrors:
SumOfSquerrors(Predic1) is 58
SumOfSquerrors(Predic2) is 45
This shows that Predic2 is "better" and we are happy :)
But what I've never understood is that if your objective is to magnify errors, why not cube it? Why not to a greater power still? If the other benefit is that all negative values to an even power become positive, then why not take the absolute value of the cube? No matter what, the degree to which we magnify errors strikes me as arbitrary.
The intuitively more logical “average absolute error” doesn’t necessarily have a unique solution. For example: if your samples are x1 and x2, any estimator between x1 and x2 has minimum average absolute error.
Squared error doesn’t have that problem (the midpoint beteren x1 and x2 uniquely minimizes it) and (very important historically) is easy to compute for the linear regression case. That’s why linear regression and squared error won. The rest, I think, is gravy. If absolute error were easy to minimize, we might even have found/invented some other properties that ‘show’ why that is nice.
Edit: turns out that article was actually featured on HN previously! See https://news.ycombinator.com/item?id=13032210 for the thorough accompanying discussion.