Suspense is a crucial ingredient of narrative fiction, engaging readers and\nmaking stories compelling. While there is a vast theoretical literature on\nsuspense, it is computationally not well understood. We compare two ways for\nmodelling suspense: surprise, a backward-looking measure of how unexpected the\ncurrent state is given the story so far; and uncertainty reduction, a\nforward-looking measure of how unexpected the continuation of the story is.\nBoth can be computed either directly over story representations or over their\nprobability distributions. We propose a hierarchical language model that\nencodes stories and computes surprise and uncertainty reduction. Evaluating\nagainst short stories annotated with human suspense judgements, we find that\nuncertainty reduction over representations is the best predictor, resulting in\nnear-human accuracy. We also show that uncertainty reduction can be used to\npredict suspenseful events in movie synopses.\n