0
$\begingroup$

I am trying to calculate some statistical parameters (min, max, average, st. dev., 90th percentile) for the difference between some observed height values and the corresponding predicted values. The difference values can be negative and positive. Should I calculate the statistics on the absolute values or leave them as they are? Why is that?

Thank you very much!

  • 0
    It depends entirely on what you are trying to measure. If you want the absolute accuracy of your predictions, so that an error of $+1$ means exactly the same to you as an error of $-1$, then take the absolute value. If you want to study the distribution of errors, however, then you'd be interested to see if, say, $+1$ came up a lot more often than $-1$. But it's up to you. – 2017-02-05
  • 0
    Just to say: as a general matter, bias in prediction tends to be interesting. If my prediction tends to be too low or too high, I'd want to know that. That means keeping track of signs. – 2017-02-05

0 Answers 0