1
$\begingroup$

I'm trying to decide whether I should scale my features & responses for training, and I'm in a situation where I can't just try both scaling and not scaling.

My features currently have an std around 0.05, and the behavior of the timeseries I'm studying is very much scale dependent (0.5 means a VERY different thing than a 0.05 in terms of what's happening in the market).

Can I expect major improvements in performance, optimizer-wise, if I scaled my features to have std of 1?

My models are different combinations of MLP and 1d conv, and my algorithm is gradient descent with the Adam optimizer.

Thank you! [and yes; I'm another one of those people who are trying to forecast the stock market]

Louis is right that I should try both. But I'm leaving the question up just in case someone comes along and gives a mathematical proof of exactly normalization is useful /useless.

$\endgroup$
  • $\begingroup$ Out of interest, why can you not just try both? Doing multiple experiments in order to find best approach is a core approach in data science, and you severely hamper your ability to do well on a task if this is not possible for you $\endgroup$ – Neil Slater Feb 9 '19 at 7:44
  • $\begingroup$ Time, mainly. Yup I'm a lazy bum. $\endgroup$ – Cedar Feb 10 '19 at 17:49
1
$\begingroup$

It's impossible to know whether the performance will improve without knowing what algorithm you are using. Even then the only way to tell is to try both.

That being said, I can't think of any scenario that standardisation would hurt the performance.

|improve this answer|||||
$\endgroup$
  • $\begingroup$ Welp I guess you're right. I'll suck it up and do both. $\endgroup$ – Cedar Feb 10 '19 at 17:49

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.