The purpose of a model is not only to maximize reward but to optimise the reward to risk ratio. Higher losses could offset that ratio even with slightly improved reward.