I want to predict the trend values of a time serie [Y] based on the effect of other 10 input variables which can also have interaction. Since the combination of interaction between the inputs is unknown, I am applying a regression Neural Network to automatically detect and take that into account. The process I follow is:

1. Get raw data of [Y] time serie per day

2. Perform decomposition and extract the [Y] Trend element per day

3. Merge observations of the [Y] Trend with the observations of the 10 variables by day into a common table

4Group the observations by unique combination of x[10] variables and average [Y] observations by "group_by" combination

5. Normalize (min-max) all variables between 0 and 1

6. Fit a NN model with:

  • Trend as 1 output node
  • The 10 variables as the input nodes
  • 1 hidden layer
  • logistic activation function

7. Perform k-fold cross-validation to estimate MAPE error

Am I following a logic approach or is it anything important that I'm missing?

  • 1
    $\begingroup$ To me it is fine, take it just as a hint that it may be fine and look at the opinion of the other members, but it should be.. the only thing I don't fully understand is why after de-trending at point 2 you re-insert the trend at point 3 among the variables and normalize it (at least this is what i roughly understood, so maybe I just misinterpreted..) $\endgroup$ – Fr1 Aug 13 at 19:49
  • 1
    $\begingroup$ Thanks @Fr1! Indeed I had to perform the detrending on an isolated table otherwise the detrend function will run across all the rest variables, including the ones I don't want to detrend. $\endgroup$ – nba2020 Aug 13 at 19:55
  • $\begingroup$ @Fri just to clarify :) at 2 I'm keeping the trend hence I'm not detrending but deseasonalizing the time serie. $\endgroup$ – nba2020 Aug 17 at 18:55
  • 1
    $\begingroup$ ok sorry I likely misinterpreted $\endgroup$ – Fr1 Aug 17 at 19:25

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.