We present surrogate regret upper bounds for online structured prediction with bandit and delayed feedback.