Applications in wireless sensor networks (WSNs) experience the routing hole problem. That is, the current node cannot forward to the destination, although it is the closest node, but not a neighbor of the destination. Accordingly, the packets from the source cannot be delivered to the destination. Jiang et al. recently proposed a SLGF approach to address this problem. However, SLGF still has long routing paths, since it uses the right-hand rule. In this paper, we describe Predictive Geographic greedy Forwarding, PGF. PGF uses information on the hole to build the virtual convex polygon, predict the routing path and choose the shorter path. PGF reduces the length and the number of hops of routing paths. Computer simulation shows our PGF scheme can reduce the average number of hops of routing paths by about 32% compared to Geographic greedy Forwarding, GF, and about 15% compared to SLGF.