There is no need to frame this as "winning versus losing" regarding the many models that we draw upon.
Even when talking about various kinds of scientific and engineering fields, predictive power isn't the only criteria, much less the best. Sometimes the simpler, less accurate models work well enough with less informational and computational cost.
Even if we focus on prediction (as opposed to say statistical inference), often people want some kind of hybrid. Perhaps a blend of satisficing with limited information, scoped action spaces, and bounded computation; i.e. good enough given the information we have to make the decisions we can actuate with some computational budget.