1
$\begingroup$

When forecasting time series one can change the problem from a classical time series (ARIMA type of models) to supervised learning (by adding lag features).

When the time series is long and you convert it to supervised learning you end up having a lot of features.

In case that we had done previous modeling with ARIMA and found the AR, I, MA, parameters.

Could we use this knowledge/parameters to select a subset of lag features?

$\endgroup$
1
  • $\begingroup$ Yes you can use $\endgroup$ Commented May 4, 2020 at 16:21

1 Answer 1

2
$\begingroup$

Yes. It could be use. if you find the lag pattern by using acf, pacf , you can generate the rolling mean or mid value time series as part of feature engineering. And use the combined feature to do supervised learn.

However, one of key important thing need note: time series based cross validate in the begining to avoid data leaking, other wise you metric will be unreasonable good.

$\endgroup$
2
  • 1
    $\begingroup$ You can, but can you guarantee (mathematically) that it will work? $\endgroup$ Commented May 7, 2020 at 6:18
  • $\begingroup$ thanks for comments and point out the weakness. I just share practice experience not mathematical based or guarantee. $\endgroup$
    – Yong Wang
    Commented May 7, 2020 at 15:52

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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