I am currently working on a dataset where the feature for each observation is time-series.

For instance assume two observations: "person X and person Y", and the feature 'price paid for milk' as time-series data for each.

Now my aim is to normalize the data as a pre-processing step and feed it into a neural network. However, for normalization, do I normalize values with respect to the time-series of each observation, or normalize across the samples?

My understanding is that the first approach induces a bias whereas the second might magnify the effect of outliers/single events due to outliers.

  • $\begingroup$ when new data comes will it be truly the price paid by new customers only? will you not get new price observations of the same people you looked at before? I find it hard to belueve $\endgroup$
    – Aksakal
    Commented May 26, 2020 at 22:15

1 Answer 1


I believe too that your series can jump all over the place with the 2nd approach, but have a (near) constant mean. That's why people detrend or difference time series.

for me, i would normalise both series in terms of each observation and compare them. To compare, you would have to apply the same normalization to both time series. And which one to use depends on correlation or variance.

Or, couldn't you index both series? Let's say the first observation in each series is 10. Then you have them on the same scale, and you can compare them


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.