Options

"I'm looking for documentation about forecasting with Neural Networks"

encripteduserencripteduser Member Posts: 4 Contributor I
edited June 2019 in Help
Hello Guys,

I'm looking for documentation about forecasting time series (sales) with neural networks, then I'll compare the performance of the NN with other classical methods like SARIMA or Exponential Smoothing.

I haven't found much content about this. Can you give me a hand?

I have the sales of 1 product that is complex to forecast (units and $), and I have more economic data like conversion rates (because imports affects the directly the sales of this product).

Many thanks in advance,
Best.

AL

Answers

  • Options
    MartinLiebigMartinLiebig Administrator, Moderator, Employee, RapidMiner Certified Analyst, RapidMiner Certified Expert, University Professor Posts: 3,517 RM Data Scientist
    Hi,
    are you sure you want to do a "direct forecast"? It is often more useful to window the document ,extract features and then run a regression or even classification model on it.

    ~Martin
    - Sr. Director Data Solutions, Altair RapidMiner -
    Dortmund, Germany
  • Options
    haddockhaddock Member Posts: 849 Maven
    Hi there,

    By complete chance I was asked a similar question recently elsewhere. SARIMA is available in R, and R is available in RapidMiner. Direct comparison therefore possible. https://onlinecourses.science.psu.edu/stat510/node/68

    Good luck!
  • Options
    encripteduserencripteduser Member Posts: 4 Contributor I
    Thanks for your answers Martin and haddock!

    I have no problems forecasting with SARIMA or ES models. My doubts are about Artificial Neural Networks.

    I've been looking applications of this models and are mainly used in classification. But that assumes a limited number of outputs.

    Question, the output of a Neural network it can be a positive variable (Y(t+h))? where h is a number of steps forward the last data available in t. This means I don't want a classification for a instance of inputs, I'd like to get a estimated number of sales for a specific time horizon.

    I'm learning about ANN and any help with this topic it will be appreciated.

  • Options
    haddockhaddock Member Posts: 849 Maven
    There is a Series extension for passing a window across data, you'd need that whatever analysis method you choose, if only to compare like with like.

    Best

    H
  • Options
    encripteduserencripteduser Member Posts: 4 Contributor I
    Cool! There is the function that I was looking for: Windowing.

    Now I'm having troubles building the input of data for the models.

    Specifically I want to do this: image


    Source: http://www.simafore.com/blog/bid/110752/Time-Series-Forecasting-from-windowing-to-predicting-in-RapidMiner

    I have my time series in the first column of an excel file, but I can't build the data set in the format required.
  • Options
    encripteduserencripteduser Member Posts: 4 Contributor I
    The problem of the last post was solved! I defined correctly the labels and the data structure went well.


    Now I'd like to know how can I get more statistics about the results of the neural networks prediction. How can I calculate the MAPE in the validation set? (assuming that I have real values and predicted values (with ANN)).

    Thanks for your help!

    Best,

    AL
  • Options
    haddockhaddock Member Posts: 849 Maven
    I seem to remember this coming up before, try searching on "MAPE".

    H
Sign In or Register to comment.