Sorry if the question seems basic. I understand that non stationary data is a big issue for traditional time series forecasting methods like ARIMA and VAR but is it the same for machine learning methods like XGBoost and neural networks? Meaning to say, do I have to do differencing on the y variable for machine learning methods like I do for ARIMA?
$\begingroup$
$\endgroup$
5
-
2$\begingroup$ ARIMA is built for non-stationary series, the I stands for "integrated" $\endgroup$– AksakalCommented Aug 18, 2022 at 2:40
-
$\begingroup$ I understand that, but the main question is - do i have to do differencing for machine learning methods like XGBoost and Vanilla neural networks for time series? $\endgroup$– oldradishsoipCommented Aug 18, 2022 at 3:03
-
$\begingroup$ You don't need to be sorry for asking a question. Just mentioned your understanding & effort about the issue along with references, if any, to show the readers you've put in enough effort before asking a new question. $\endgroup$– Dr Nisha AroraCommented Aug 18, 2022 at 6:21
-
$\begingroup$ I've tried finding googling for answers but the views are always contradictory. Some say it's better to do differencing whenever possible but there are some who say that detrending the y variable is not ideal especially if your x variable is not detrended as well $\endgroup$– oldradishsoipCommented Aug 18, 2022 at 6:41
-
$\begingroup$ The short answer is yes, you need to difference to preserve neurons and limit computations $\endgroup$– AksakalCommented Aug 18, 2022 at 10:47
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
1
No, you don't have to do differencing when using XGBoost or neural networks. The idea with those data-driven models is that they have so many parameters that they can adapt to any property of the data. You don't have to presume any model, meaning you don't have to do the work to figure out an appropriate model. The disadvantage of data-driven models is that you learn much less about the "inner properties" of your data-generating system. But if you only care about prediction, that is totally fine.
-
$\begingroup$ Thank you for the answer. As far as I understand, we're talking about non-stationarity in the target variable. Would your answer change if we have one or several non-stationary predictors with, say, a binary target in a panel data? $\endgroup$ Commented Jul 24, 2023 at 14:27