How do I autoregress on regressors that aren't the target?

[tszumowski]

I noticed the AutoRegressivePipeline autoregresses on the target variable to build out the multi-step forecast output. I have a variable that I lag alongside the target, because it is a useful regressor, but it is not something known at the time of forecast.

For example, consider the case where trying to predict a store's product sales by week. The target is sales. However, we may have a regressor, customer_visits that indicates the number of times a customer walks in the store. If I am forecasting for the next 4 weeks, I do not know the values of customer_visits ahead of time. But I can forecast it alongside sales and use it to forecast sales.

Is there a way to do this? If not, is there a suggested alternative?

[Mr-Geekman]

I can see two ways here:

1. You can use lagged version of the customer_visits with lag >= 4. In this case it will be available on the 4-th week.
2. You can try to build a separate model for prediction of customer visits on test data and then feed its predictions to model, that treats customer_visits as regressor. There is a problem with this approach: on train data we use true values of customer_visits, but on test data we use predicted values that are not precise but model will rely on it like it is precise.