0
$\begingroup$

I have the following case for which I am needing to forecast a value, say, 12 months out:

  • Many individual entities each with their own time series. Each entity has the same data structure
  • For each series, say 24 - 36 months of historical data exists (each entity is "created" and then "matures")
  • Each entity will mature differently (and as of yet, unpredictably) so we'll use multivariate training data to hopefully capture that unknown signal in suspect variables
  • We are likely only interested in that 12 month step forecast, and possibly only once (made early on, like 6 months in)
  • Forecasts are to be made on new entities based on learning from the old

I'm a bit at a loss as to whether this is better modelled with something like ARIMA/time-series like approaches (which I am less familiar with) or simply multilinear regression with some lagged columns. It seems a mix of the two to me, but is there any way in which either is more appropriate here?

$\endgroup$
4
  • $\begingroup$ Vector autoregression (var) and panel data model (panel-data) come to mind. $\endgroup$ Apr 10, 2021 at 14:10
  • $\begingroup$ Indeed it seems like I would maybe be contrasting a VARIMAX type implementation against a simple linear regression with lags possibly (and maybe multiple models if multiple prediction steps are needed), but my knowledge stops there $\endgroup$
    – Josh
    Apr 10, 2021 at 15:23
  • $\begingroup$ Are you trying to predict one variable with another? Or do a univariate prediction. Since you are talking about var models I would guess the former, although it is not clear to me from the post what you are trying to predict with what. You might also consider autoregressive distributed lag models. To me VAR models are very hard to interpret. :) And you can not have many predictors in practice. But I would take Richard Hardy's advice I am hardly an expert. :) $\endgroup$
    – user54285
    Apr 13, 2021 at 2:38
  • $\begingroup$ I mention in my post it's a multivariate problem -- I'm predicting the state, call it y, at a certain time using multiple variables (including the prior values of y that are available when the prediction is made). Interpretability is not an issue here (I would be fine with a black box model). $\endgroup$
    – Josh
    Apr 13, 2021 at 23:54

0

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.