Before training the model, I convert a target value to the log scale since range of the target is quite large.

After training the model, the Absolute Mean error was estimated as, for instance, 0.5.

If I want to re-convert the predicted target value to linear scale, how can I handle the error on the target value?

For example:

A predicted target value = 2.5 (at log scale)
The Absolute Mean error = 0.5 (at log scale)

When re-converting the predicted target value to the linear scale while including the error, would the range of the target value be exp(2.5-0.5) ~ exp(2.5+0.5)?


1 Answer 1


I think this is incorrect.

If on the log scale you have log(value) + log(error), in order to get back something of the form value + error onthe linear scale, you would need to use these exponents separately, in other words value + error = exp(log(value)) + exp((log(error)).

With your method, you get exp(log(value) + log(error)) = exp(log(value))*exp(log(error))=value*error which is not the desired value + error.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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