Apologies if the question is not fit properly here.

Goal: I have a system identification project where I need a dataset to construct a dynamic model and perform correlation and/or predictions. The methods are linear (AR,ARX,ARMAX,ARMA,Kalman,nonparametric methods) because I have not learned nonlinear methods in my course.

Question: I wish to ask if there are available datasets (downloadable as csv files) that are uploaded by users or authors of articles in the field of:

  • Environmental Modelling: For instance, creating a time-series forecast model for temperature vs forest fire using AR/ARMAX...
  • Biomedical Modelling: For instance, creating an ARX model for blood pressure flow...

Issue: I have scanned the entire web searching for datasets to construct system id models but the issues I am facing are:

  1. Most articles use Nonlinear methods for cited datasets (such as neural networks).
  2. Most of the articles that use linear methods are in the early 2000s and the datasets they cite are no where to be found.
  3. Most datasets where system ID is used are in a domain (civil/mechanical/chemical engineering) different than mine which is in environmental modelling or biomedical modelling.

Therefore I would kindly ask for your help in referencing me to any dataset that could meet the two requirements above.

  • $\begingroup$ Do you need a reference of the actual system or just data? $\endgroup$
    – Royi
    Apr 17 at 8:22
  • $\begingroup$ I would like reference for the data even if no literature involved in that data exist. As long as linear system ID works. @Royi $\endgroup$
    – SPARSE
    Apr 17 at 8:26
  • $\begingroup$ I am not sure what do you mean by ID. I meant do you need data or data + labels (The system which generated it)? $\endgroup$
    – Royi
    Apr 17 at 8:30
  • $\begingroup$ Oh, I would prefer data+labels in this case @Royi $\endgroup$
    – SPARSE
    Apr 17 at 8:33
  • 2
    $\begingroup$ Idea: pick only samples on which linear methods work well? It's what others would have to do anyway, for such a dataset to exist. Aim for 90%, and explain why linear methods work there and not for the other 10%. $\endgroup$ Apr 18 at 21:15

1 Answer 1


Such datasets either don't exist, or cheat. If linear methods could achieve 90% on real world forecasting tasks, we'd need not sink billions into neural nets and the like.

An option is to create such a dataset yourself, by picking samples such that 90% is achieved - it's what others would have to do anyway. Then, try to explain why the linear methods work there but not on the other 10%. Another option is to apply a nonlinear (but interpretable) preprocessing step ("feature extraction") then apply linear methods.

Fundamentally, most real-world structures aren't linearly separable. Nonlinearities enable such separation, and hence classification and regression. Example - nonlinear projection into a higher dimension, enabling linear separation (with a hyperplane):


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.