I am facing an issue with a time series prediction problem. My data looks like the following:

Datetime , Feature 1, Feature 2, Feature 3, Feature 4

19-03-2015 02:15 ,80 ,50 ,16 ,0 ,4

19-03-2015 02:40 ,80 ,3 ,50 ,2 ,1

20-03-2015 15:40 ,54 ,8 ,0 ,5 ,3

............for 2 years As you can see the input data has unevenly spaced time intervals, and additonmally the sum of features 2,3 and 4 is less than or equal to the value of feature 2. So there is depedency amongst features as well. Plus the data is given for 2 years(2015, and 2016), and there might be trends/seasonality in there as well. Now the challenge is to predict the Feature 2-4 on the test data set, which has provided only the Date, and Feature Values for the year 2017 ( again at arbitary time intervals). For example:

Datetime , Feature 1, Feature 2, Feature 3, Feature 4 19-03-2017 12:30 62
21-03-2017 22:00 70
21-03-2017 15:40 60


  1. this problem has not only unevenly spaced intervals in the input
  2. Possible seasonaliy and depency constraints between features
  3. possinble seasonality
  4. possibly Three different parallel time series? for feature2-4

I thought of making a single training example as descrontructing the date/time variable into 3 features, weekday, hour, minute. Then adding it to values the Feature 1-4, thus totalling 7 features for time step t. Then including the 3 date features from the next timestep t+1, as well as feature 1 from t+1. The output being Feature values, 2,3,4 for timestep t+1. But this doesnt take the possibility of three different time series..

Is this a valid approach? Or does anyone have any suggestions or LSTM code for their approaches?


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.