Since we're talking about probabilistic decision theories, if you consistently make "good decisions" you will still obtain "bad outcomes" some of the time. This should not be cause to start doubting your decision procedure. If you say you are 90% confident, you should be thrilled if you are wrong 10% of the time - it means you're perfectly calibrated.

A perfectly rational agent working with incomplete or incorrect information will lose some of the time. The decisions of the agent are still optimal from the agent's frame of reference.