My question is about the intuition for hyperparameter tuning of time series.

In other models, like Linear or Logistic Regression there is labeled data and according to accuracy or precision, the parameters are tuned. But in time series, the future values are predicted. For example, the next 30 days.

My question is, what is the point of tuning, when you do not know what the answer is (as it is going to happen in future)?

  • $\begingroup$ In this respect time series data does not differ from tabular data. You can simply use the values for which you know the label and use part of that data in a test dataset to tune your hyperparameters. $\endgroup$
    – Oxbowerce
    Dec 22, 2021 at 18:12
  • $\begingroup$ @Oxbowerce so you mean, if I have data from 2010 to 2014, I can tune parameters on 2013-2014 "predictions". After that, with tuned parameters I can predict 2015? $\endgroup$
    – Anar
    Dec 22, 2021 at 19:02
  • $\begingroup$ In you example you could use 2010-2013 to train your model, use 2013-2014 to see how the model performs on unseen data (since this is still historic data you have the labels/actual values for that period so you can calculate the performance metrics) and then use the model to start predicting new values for the future. $\endgroup$
    – Oxbowerce
    Dec 22, 2021 at 20:25


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.