{"id":40950,"date":"2026-05-13T10:27:09","date_gmt":"2026-05-13T08:27:09","guid":{"rendered":"https:\/\/www.graviton.at\/letterswaplibrary\/how-to-apply-normalization-for-cross-sectional-time-series-data\/"},"modified":"2026-05-13T10:27:09","modified_gmt":"2026-05-13T08:27:09","slug":"how-to-apply-normalization-for-cross-sectional-time-series-data","status":"publish","type":"post","link":"https:\/\/www.graviton.at\/letterswaplibrary\/how-to-apply-normalization-for-cross-sectional-time-series-data\/","title":{"rendered":"How To Apply Normalization For Cross Sectional Time Series Data ?"},"content":{"rendered":"

<\/p>\n

\n

I am unable to convince myself to use one method.
Some methods that i thought of were :<\/p>\n

    \n
  1. I use normalization for full training data of one subject across all features. In this method, i am introducing some kind of lookahead bias, and also this loses on some information which could have been valuable. And also when i want to use one model ( suppose regression with gradient descent) for the subjects combined, then I am unable to judge if this will be a good method.<\/li>\n
  2. A bad method was to not care about the subjects, and just normalize across full feature. but this just feels wrong to me.<\/li>\n
  3. I was reading about cross sectional normalization which ranks the subjects and does some kind of normalization. But i am unsure how that would be useful.<\/li>\n
  4. Another way i found was by using some rolling window, where i keep normalizing not over full data, but the past window data. This seems better but here also what choice of window should be done, and there are lot of questions.<\/li>\n<\/ol>\n

    And the bigger problem over all of these is the time series . I would lose quite a lot of information when i don’t consider these. ( although not all features have a big factor of this).<\/p>\n<\/div>\n

    submitted by \/u\/Virtual-Current6295 <\/a>
    [link]<\/a><\/span> [comments]<\/a><\/span><\/p>