Good question.  The statistical and dynamical models shown in the popular IRI/CPC plume of models do not all have the same "initial condtions," meaning they do not all have the exact same input data.  Dynamical models often have a huge amount of atmospheric and ocean data, while the statistical models only receive input from less than a handful of variables (most often sea surface temperature).   Also, the input data into these models can be based upon different time averages.  In particular the statistical models often rely on monthly or seasonal (3-month) averaged inputs, while the dynamical models generally rely on daily or shorter averages.  So we tend to believe the dynamical models are at an advantage simply because they see the most recent observational evolution better than many of the statistical models.  However, there are certainly cases where statistical models performed better, so clearly this is only part of the story.  But overall, in the last 10 years, we have generally seen better model performance from the dyanamical models over the statistical.  Here is one paper that explores that:

http://journals.ametsoc.org/doi/full/10.1175/BAMS-D-11-00111.1 (is free/open access)