# Questions concerning Z-Normalization in Dynamic Time Warping

Here I found this very nice presentation. On page 46 one can read the following:

Essentially all datasets must have every subsequence z-normalized. There are a handful of occasions where it does not make sense to z-normalize, but in those cases, DTW probably does not make sense either.

I tried to use Z-Normalization with the following code in Python:

trainX = stats.zscore(trainX)
testX = stats.zscore(testX)


I.e. I normalized the entire data-set. This resulted in a slighlty worse performance of my classifier. On page 48 of the presentation mentioned above I read the following:

Preempting a common misunderstanding: It is not sufficient to normalize the entire time series. You must normalize each subsequence.

Concerning this I have the following questions:

1. Does the above statement mean that I should normalize the entire time series and each subsequence or just each subsequence?
2. What's the reasoning for either choice?