The grid values should not be considered as representing the weather conditions at the exact location of the grid point.  They should be considered as a time-space average within a two- or three-dimensional grid box.  The discrepancy between the forecast grid-point value and the verifying observed average value can be both systematic and non-systematic:

 

Fig3.2-1:  Comparison between NWP model output and observations ought ideally to follow a two-step procedure:


Fig3.2-2:  In reality, the comparison between NWP and observations must for simplicity bypass the area average stage.  This results in the systematic and non-systematic errors arising from distinctly different sources.  The effects related to the two green arrows in Fig3.2-1 are here combined into one.


Systematic errors maybe due to model deficiencies and/or observational representativeness.  These can be partly corrected by statistical means (e.g. model output statistics MOS).  A series of forecasts also helps with dealing with uncertainty.

Non-systematic synoptic errors can be dampened by different ensemble approaches (e.g. medium range ensemble, probability considerations, forecast error growth).

Sub-grid variability (notably for rainfall but other parameters too) can be addressed by downscaling.  Downscaling converts the grid box area average probability density functions from the raw ENS into "point rainfall probability density functions" for points within each grid box.

New downscaling techniques are being developed accordingly (see for example the Point Rainfall product).