Prediction by partial match (PPM) is a commonly used technique in Web prefetching, where prefetching decisions are made based on historical URLs in a dynamically maintained Markov prediction tree. Existing approaches either widely store the URL nodes by building the tree with a fixed height in each branch, or only store the branches with frequently accessed URLs. Building the popularity information into the Markov prediction tree, we propose a new prefetching model, called popularity-based PPM. In this model, the tree is dynamically updated with a variable height in each set of branches where a popular URL can lead a set of long branches, and a less popular document leads a set of short ones. Since majority root nodes are popular URLs in our approach, the space allocation for storing nodes are effectively utilized. We have also included two additional optimizations in this model: (1) directly linking a root node to duplicated popular nodes in a surfing path to give popular URLs more considerations for prefetching; and (2) making a space optimization after the tree is built to further remove less popular nodes. Our trace-driven simulation results comparatively show a significant space reduction and an improved prediction accuracy of the proposed prefetching technique.