On Thursday, 25 February 2016 13:49:56 UTC, wrote:

Data assimilation is about nudging the model to a solution that better fits
the observational data. It does need actuals to fine tune its analysis.
After all forecasting short term is basically an initial value problem, and
so if data assimilation is done every 3 or 6 hours then nudging is required.
When continuous assimilation comes onboard every piece of data will be vital
since the short time-scales will require a greater accuracy on spatial
scales to avoid numerical instability in the assimilation process.

There must be myriad examples where actuals are vital for the short-range - the thundery low coming up from the south where rainfall totals and location of storms is difficult or knife-edge rain/snow situations where half a degree of lowering in heavy rain turns it to snow etc.

Richard