On the predictive performance of a non-optimal action in hypothesis testing