We should introduce tags for the following properties of forecasters:

• can make multivariate forecasts
• can make use of exogeneous data (doesn't just ignore it)
• can deal with non-equally spaced time series
• assumes that time stamps of exogeneous data are a super-set of endogeneous data
• can deal with variables sampled at different time stamps
• can deal with exogeneous data that are not known throughout the future

All of these properties should map on input checks which raise informative errors.

We should also think of standard adapters that turn the restrictive version of the tag in the non-restrictive one.

For example:

• a wrapper that turns algorithms into multivariate versions by iterating over variables in y. There are two options here: consider the other y variables as exogeneous or not (latter should be the default).
• a wrapper that uses an interpolator to reduce to the equally spaced case; this needs to have transform (re-sampling equally) and inverse transform (interpolating between equally spaced forecasts)
• removing the requirement to see fh in fit by moving fit into predict at the price of inefficiency