You must decide that.
There are methods to interpolate the NaN values if you want to do that, one of which is the fillmissing function (R2016b and later).
It would likely be best to do the forecast with and without the NaN values (some functions omit them automatically) and see if interpolating them affects the forecast.