2
$\begingroup$

Usually, the validation error is higher than training error, but are there any cases when they are equal?

$\endgroup$
1
  • $\begingroup$ I have seen cases where the test error is lower than train error. Model is a gbm and data is of sequential nature. $\endgroup$ – horaceT Oct 3 '16 at 1:41
3
$\begingroup$
  • Reason 1: the model is underfitted, i.e. it has a high bias:

enter image description here

  • Reason 2: the model is near perfect.

  • Reason 3: the training set is very similar to the validation set, e.g. if some data from the validation set have leaked into the training set:

enter image description here

  • Reason 4: if using a neural network, the training have been prematurely stopped:

enter image description here

$\endgroup$
0
$\begingroup$

Ideal, anything can happen depends on the distribution of training data and testing data. I can easily craft an example that make training error and testing error to any value you want.

In real world, if you build model very carefully and you have large amount of data. The testing error should be very close to training error. This is because we are trying to use math to describe a physical relationship. If the model is good enough, it should be applicable in future.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.