I'm using ARIMA model for time series forecast. My data has increasing variance and I applied a BOX-COX transformation to stabilise it. Here are charts: enter image description here

After I run my app it turned out that ARIMA produces slightly better accuracy for raw data:

  • Raw data MAE error: 4.83
  • BOX-COX transformed data MAE error: 4.97

As I understand ARIMA should produce better results for stationary data. Base on that I should see better accuracy when I'm using BOX-COX transformed data right?

I'm wondering why in my case model gives worst results for stationary data and what can I do to improve my model?

  • 1
    $\begingroup$ is this financial returns data? $\endgroup$ – Taylor Feb 17 '19 at 2:36
  • 1
    $\begingroup$ Per @Taylor's comment, if these are financial returns, then explicitly modeling the variance through ARCH/GARCH is probably better. Also, did you adjust for bias in back-transforming? $\endgroup$ – Stephan Kolassa Feb 17 '19 at 6:17
  • $\begingroup$ @Stephan Kolassa, ARCH/GRACH is one of models what I will definitely use. I just started learning about forecasting/ML and I'm curious why ARIMA with stationary data produces bigger error. I didn't adjust bias. THX for tip! $\endgroup$ – Sayaki Feb 17 '19 at 20:08
  • $\begingroup$ @Taylor, yes this is financial data. $\endgroup$ – Sayaki Feb 17 '19 at 20:08
  • $\begingroup$ looks like bitcoin or something. Sort of difficult to model the whole window because the scale process doesn't look stationary $\endgroup$ – Taylor Feb 17 '19 at 20:36

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.