Adding binary regressors
The first thing to consider with additional regressors, whether binary or continuous, is that you must have known future values for your entire forecast period. This isn't a problem with holidays because we know exactly when each future holiday will occur. All future values must either be known, as with holidays, or must have been forecast themselves separately. You must be careful though when building a forecast using data that itself has been forecast: the error in the first forecast will compound the error in the second forecast, and the errors will continuously pile up.
If one variable is much easier to forecast than another, however, then this may be a case where these stacked forecasts do make sense. A hierarchical time series is an example case where this may be useful: you may find good results by forecasting the more reliable daily values of one time series, for instance, and using those values to forecast hourly values of another time series...
