Can I assume the normality of the prediction interval for an arbitrary machine learning model? Or in general, the prediction error can follow any distribution?
$\begingroup$
$\endgroup$
2
-
2$\begingroup$ You can assume whatever you want, no matter how incorrect it would be. But if the question is if using normal distribution for prediction intervals is always correct, the answer is obviously not. The simplest example is when you predict probabilities and use normal prediction intervals that can be in $(-\infty, \infty)$... $\endgroup$– Tim ♦Feb 20, 2019 at 9:21
-
$\begingroup$ Often (but of course not always) on some suitably transformed version of the target quantity. Without understanding the problem at least somewhat it will be hard to guess a good transformation. $\endgroup$– BjörnFeb 20, 2019 at 9:37
Add a comment
|