I have an ARIMAX model that has AR & MA terms along with exogeneous variables at prior lags. I am predicting on a continuous process and attempting to forecast to 1 and 10 units ahead. However, the important lags for the exogeneous variables are within the 1-10 units. Thus, it fails to produce 10 units ahead forecast on the most recent data.

For example, I have a x variable with a lag of 3 (x_3) in the model. I just received the most recent result (0 units ahead).

If I forecast to 10 units ahead, that forecast requires the x_3 to be non-missing at 7 units, which isn't possible since it hasn't occurred yet. Therefore, no prediction is made at 10 units ahead.

How do I resolve this issue? Do I attempt to fit an ARIMA model to the x variables? Use the last value for x and propagate it forward? Am I using this method incorrectly? Create a new model that only uses exogeneous lags that exceed the forecast step size (x_11 for the 10 step forecast)?

  • $\begingroup$ If you use forecasted values for input series then make sure your software implementation incorporates that uncertainty into the confidence limits for the outpur series $\endgroup$
    – IrishStat
    Commented Nov 29, 2022 at 17:08

1 Answer 1


You will need to provide "future values" of the exogeneous variables. These can be "known" values, e.g., if they are promotions in a retail context. Or they could be "forecasted" values, e.g., if they are weather data. You could also try multiple different values, e.g., for scenario analysis.

Yes, especially the latter means that ARIMAX or regressions with ARIMA errors (there is a difference) are dubious the farther out we forecast.

This holds regardless of whether you use lagged or non-lagged values of your predictors. What you don't know you will need to forecast or assume.


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.