There is nothing you can do to predict a random process except always predict the most frequent outcome. These are known as bandit problems. When added in as noise to another process uses the mean value and beware of dangerous outliers.