A common practice in Machine/statistical Learning is to split up the dataset into a test and training set. However, 2 of my textbooks on time series analysis never do this, they apply the TS algorithms on the entire dataset to predict the next datapoint in the series.

Does anyone know why this is the case? I.e, the reason for the difference in approaches?

  • $\begingroup$ Take a look: blogs.sas.com/content/forecasting/2016/03/18/… $\endgroup$ – Zen Aug 19 '17 at 21:32
  • $\begingroup$ There are different assumptions of data in time series. Order matters, so you have to be careful on how you split your data. However, just a simple train/test split is possible and often done for time series. $\endgroup$ – Jon Aug 20 '17 at 4:51
  • $\begingroup$ Hi. But, won't this include bias in the model? I.e., "making a choice" on where / what time to split the data? $\endgroup$ – Thomas Moore Aug 20 '17 at 23:27

When you split the data and do model identification the resulting model may depend on the point of the split. If you then measure performance two things are in play the form/parameters of the model and future values. Partitioning data into "training" and "non-training" requires the assumption of a fixed model with a set of fixed coefficients/parameters.

Modern time series analysis ( probably more modern than your textbook as textbooks are out of date before they are printed !) would suggest forming a model based upon the total # of data points AND then testing for parameter divergence/change possibly suggesting then that the most recent homogeneous data be used to form a useful model.

| cite | improve this answer | |

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.