Right. In cases like this, you either a) have unmodeled dynamics or b) (much less likely) you found a superior strategy.

If b), try it out in the wild. If a), add assumptions that make the observed behavior optimal again, and see if those assumptions apply to the real world situation.