A network device in a network edge receives content directed to a mobile device attached to the network edge from an upstream network device and forwards the content toward the mobile device. The network device makes a decision whether to cache the content based at least in part on a popularity of the content in a region covered by the network device and a prediction error for an estimated probability that the mobile device will transition from the region to another region. The popularity is directly correlated with a first bias toward caching the content. The prediction error is inversely correlated with a second bias toward caching the content. The decision is implemented: the network device either caches the content or foregoes caching the content, in accordance with the decision.