5
$\begingroup$

I have a dataset, where target value changes in time in following way:

enter image description here

I need to predict target value for upcoming month, however I struggle to find a method to extrapolate the function that defines target. It is certainly not linear.

I tried to extrapolate the max/min peaks separately using linear regression, but not sure what to do with all other values.

$\endgroup$
5
  • $\begingroup$ If it were that same kind of pattern but without the increasing amplitude, what would you do? $\endgroup$
    – Dave
    Jan 25 '20 at 15:24
  • $\begingroup$ I would calculate the period of this pattern and predict the same values based on position inside the pattern. $\endgroup$
    – Eldar
    Jan 25 '20 at 18:35
  • $\begingroup$ I would say to do something similar but calculate some kind of stretching constant. The amplitude seems to be increasing linearly. $\endgroup$
    – Dave
    Jan 26 '20 at 1:51
  • $\begingroup$ Thanks, that seems like a reasonable thing to do. Do you have any suggestions about the models that can handle such constant? $\endgroup$
    – Eldar
    Jan 26 '20 at 20:14
  • $\begingroup$ I would do something with how the high peaks increase and how the low peaks decrease and use those seemingly linear trends to say how the shape is getting stretched up and down. The stretching up is more pronounced than the stretching down. $\endgroup$
    – Dave
    Jan 26 '20 at 21:30
1
$\begingroup$

That function can be decomposed into composite components:

  • General trend
  • Repeated frequencies (looks like there are 3 major cycles)
  • Random noise

A Fourier transform can decompose a function into its constituent frequencies.

$\endgroup$

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.