Improve spatio temporal prediction tests about timeseriesprediction.jl HOT 3 CLOSED

1. If you wish, but it will need to be a quick test. Tests take long enough already...

2. I do not understand

3. I do not see the point for maximum errors.
   I have seen lots of these predictions and average errors are predictable.
   Only for very long training sets ( not possible on travis ) the maximum errors go down a little bit and a test
   @test 0 < predicted_value < maximum(field) does not seem particularly useful

from timeseriesprediction.jl.

If you wish, but it will need to be a quick test. Tests take long enough already...
Travis has a time limit of 50 minutes, we are waaaay below that, but I understand your point.

But yeah in general one more of these tests is only about 1 minute. I believe we can afford this.

I do not understand
So that we are sure that the actual scale of a field is preserved. As I said in chat, the field scale is an invariant of the method. So: compare the maximum and minimum value between prediction and reality, and make sure they are "close enough" (relative scale: 0.1).

I do not see the point for maximum errors.

Okay, for this you are right, but minimum should be tested and it is easy: relative minimum error should be between 0.0 and 0.1

Relative error = abs_error / 2*(maximum field value - minimum field value)

from timeseriesprediction.jl.

ok. Fair enough.
The scale of barkley is pretty close to 0-1 anyway. So I will just add minimum(abs_err) < 0.1.

from timeseriesprediction.jl.