I have time series GDP growth rate data that I use as my Y and other X variables that I put into neural networks to make predictions. The two questions that I have are:

  1. When I decompose my GDP_growth variable I get a detrended variable that I model and make predictions of. If I have train and test data until 2023 and want to make predictions for 2024 them they will likely be way different than the real values as when I removed the trend term I got smaller numbers for the new y. So how to deal with this? When I get predictions of 0.0034 because the data it is fed (after it being decomposed) is between -0.05 and 0.05 interpretability is impossible as the real values are generally bigger than 0.05 and smaller than -0.05. This is not the networks fault as this is all the data that it sees still the interpretability is awful.

  2. When I scale my X's and put new data in (for example the year is 2025 and want to use the same model trained up to 2023 to make predictions) should I train a scaler upto 2023 and apply it to the 2025 data? If so, any unseen data for 2025 will be above the max or below the min of the min-max scaler.

  • 1
    $\begingroup$ So how to deal with this? How about adding the trend?! If you model y = f(t) + bx + e, where f(t) is some time trend then obviously bx is good prediction of y - f(t) = bx + e, but to get a prediction of values of y from bx you would have to add the trend f(t). Maybe I misunderstand you ... $\endgroup$ Commented Mar 3 at 16:06


Your Answer

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