"Long term time series prediction"

LukasLukas Member Posts: 2 Contributor I
edited May 2019 in Help
I have created a model to predict the next value of a time series from three previous values (now real data for now, just trying). It predicts the training data well, but I would like it to go further: it should append the predicted value to the original ones and use it to create another prediction window (predict based on predicted values).

window        |    prediction 
1    2    3    |    4p
2    3    4p  |    5p
3    4p  5p  |    6p
4p  5p  6p  |    7p
How do I do this in RapidMiner if it is possible at all? Are there other options for long term time series prediction?


  • landland RapidMiner Certified Analyst, RapidMiner Certified Expert, Member Posts: 2,531 Unicorn
    I would rather recommend to build a model for each horizon you are planning to predict. With this way, you can evaluate each single model and get a reliable estimation of the accuracy of the model. Otherwise the model errors will just multiply and you won't even notice it.

  • vtsimoesvtsimoes Member Posts: 2 Contributor I
    I've the same problem, and search for solution about 3 or 4 days and don't achive one. Someone help me please, this should be simple, but not in RapidMiner.
  • MariusHelfMariusHelf RapidMiner Certified Expert, Member Posts: 1,869 Unicorn
    Hi, did you install the Time Series extension? You can install it via "Update RapidMiner" in the Help menu. In this extension you find the Windowing operator, which can be used to tackle the above problem.

  • vtsimoesvtsimoes Member Posts: 2 Contributor I
    Ok, I've installed the Time Series plug in, and operator Windoing is used to generate a table off-line, but I need in run time execution generate this table, i.e., each value predict under the model, is used as a past value on the serie to predict another one, in the next step. Do you understand?
  • MariusHelfMariusHelf RapidMiner Certified Expert, Member Posts: 1,869 Unicorn
    As Sebastian posted above, it is probably a better idea to use the horizon parameter and train a separate model for each point in the future. The reasons are also stated in Sebastian's post.

    If you think that you really need your setup as you described it, probably a bulky construction with loops and some preprocessing operators might work.
Sign In or Register to comment.