Virtual Machine (VM) migration is an important technology to support Infrastructure as a Service (IaaS). Traditional pre-copy and post-copy strategies could function well in LAN but will need considerable time to migrate between remote hosts in WAN. In this paper, we propose a prediction-based strategy to optimize cloud VM migration process over WAN. In this strategy, information about size increments of snapshots is used to determine appropriate time points for migration in order to reduce the downtime during migration. Specifically, we utilize Markov Chain Model to predict the future increasing speed of snapshots. The experiments on KVM showed our approach could achieve satisfying results.