These are plots of some of my features and rest of the others having similar pattern. The data here is spanned for 2 days and I need to predict labels for 3rd day. It has 60,000 samples (seconds).

Any popular time series regression methods or repos I must be aware of to solve this kind of problems? I don’t need to forecast as I have labels for validation.

Also what are current trends for statistical models vs ML models.

Does considering lag or sliding window the only popular and effective option?

https://preview.redd.it/9vjh75m7ki2c1.png?width=1354&format=png&auto=webp&s=f1da6b7fcda0cd96f4b997b818b0272f911cf120

  • house_lite@alien.topBEnglish
    1·
    1 year ago

    R AutoQuant has ML based models that handle single series and panel data with lots of feature engineering options.

  • trial_and_err@alien.topBEnglish
    1·
    1 year ago

    If the oscillations are periodic then Fourier series can capture them. If they’re white noise you can’t predict anything. If they’re autoregressive white noise you can fit an ARMA model on them.

  • eamonnkeogh@alien.topBEnglish
    1·
    1 year ago

    Based on this figure, simple persistence would work very well.

    Persistence says, the predicted value, is the same as 24 hours ago.

    If the data has cultural effects, then the predicted value, is the same as 7 days ago (because, a Saturday is not like a Friday).

    • ade17_in@alien.topOPBEnglish
      1·
      1 year ago

      100%. I did exactly the way you described and it works awesome. And got to know it’s called “persistence”.

      I implemented a very interesting and creative approach to train such a model, and will surely make a post here later for a detailed summary.