Is z-normalization of time series subsequences a good idea?

[seanlaw]

Often, time series are de-trended or pre-processed before any analysis. With matrix profiles, this is overcome by z-normalizing each subsequence before computing the distance with respect to another subsequence. The question is whether or not z-normalization is a good idea?

[seanlaw]

According to Eammon Keogh:

Z-normalization means we are ignoring the mean and standard deviation of the data. At least 99.9% of the time, that is the right thing to do.

Furthermore, this 2023 research paper by Felipe Tomazelli Lima and Vinicius M.A. Souza also seems to support this claim with the caveat that:

alternative methods can outperform z-normalization in the classification task

[JaKasb]

Search for similar system-behavior, (sequences of up&down sections) : z-normalization
Search for similar magnitude or energy : no-normalization

As alternative to mean-normalization : for preprocessing the time-series apply a high-pass fiter for removing slow drifts and varying bias-values.
E.g. window-size = 1 second -> lowest frequency in pattern = 1Hz -> Set high pass filter frequency below 1Hz to remove drift.

[NimaSarajpoor]

Search for similar system-behavior, (sequences of up&down sections) : z-normalization

Search for similar magnitude or energy : no-normalization

@JaKasb
Correct. This is similar to what Eamonn said before. You cannot have both shape and magnitude. When shape matters, the magnitude does not, and vice versa.