1
$\begingroup$

I have 1000+ different timeseries. I want to build one ARIMA model which gives better forecast for most of the timeseries. I went through the answer of this question Estimating same model over multiple time series

The answer given my rob hyndman caught my eye. Can someone explain that answer? what padding is he talking about? how to do padding for weekly data and how to do it in python?

Thank you

$\endgroup$
1
$\begingroup$

I believe that he is combining all of the time series into 1 long time series. Then he is padding the periods in between to make sure the the time series line up (i.e. each January for each time series actually occurs in January) which could be an issue if your time series are of different length or the same length but not full years. Also he doesn't want any serial correlation between the series so he adds several periods to ensure that.

As others have pointed out in the comments this assumes that the overall level of the each time series is the same. For your case with 1000 time series I would probably use the accepted answer of the thread and just say there is an 'on average' better arima model.

I think Hyndman's approach could be a nightmare with a large data set like yours.

$\endgroup$
1
  • $\begingroup$ Yes, if the combined time series becomes quite long, estimation might become practically infeasible. Maybe still feasible for an AR model (based on OLS), but not ARMA (maximum likelihood via state space). $\endgroup$ – Richard Hardy Jun 14 at 20:32

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.