It depends on if the problem is convex or non-convex. If the problem is convex, (considering a well-scaled, well-bounded model), using the "crossover" stop criterion should yield the global optimum. But, if the problem is non-convex this criterion can stop prematurely. If this is the case, the (heuristic) NLP worsening criterion is advised to be used, which of course doesn't GUARANTEE that the optimum is global. ( I'm just repeating some of

https://www.gams.com/latest/docs/S_DICOPT.html )

From my experience, the most important considerations for DICOPT to work properly are modeling ones (good scaling, good bounding). Of course, you should always use a global solver (or whatever procedure) once to get the global optimum, use this to tweak DICOPT and then enjoy its improved performance.

Best

Claudio