4
$\begingroup$

I'm working on a time series problem, with additional predictors. While I'm exploring various ways to approach the problem, one possible way is to turn the time series problem into a supervised machine learning problem, i.e. use $(t-k), ..., (t-1)$ to predict $t$, where $k$ is the number of lagged values (as described here). I also plan to use additional predictor (binary variables) such as holidays, days of the week, months of the year, etc.

For now, I need to determine variable importance of each predictor (including lagged values). Is it valid to use Random Forests to determine variable importance in such a problem? If not, what are other (better) ways to approach the problem?

Note that I understand that there are methods specifically for investigating the correlation of a $t$ and its lagged values $(t-k), ..., (t-1)$ (e.g. ACF, PACF). However, I'm also interested in the importance of the exogenous variables.

$\endgroup$
2
  • $\begingroup$ Whether this will be be useful almost certainly depends on what the variable importance measure is meant to indicate. What do you want to use it for? Since you mentioned more traditional time series analysis methods, note that there are also vector autoregressive models with exogenous inputs $\endgroup$
    – user20160
    Commented Apr 6, 2018 at 4:07
  • $\begingroup$ I think it would be insightful in feature engineering, i.e. understanding if a feature will be valuable in the the forecasting task. $\endgroup$
    – meraxes
    Commented Apr 10, 2018 at 5:49

0

Your Answer

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