Phylogenetic tree reconstruction is difficult in the presence of lateral gene transfer and other processes generating conflicting signals. We develop a new approach to this problem using ideas borrowed from algorithmic information theory. It selects the hypothesis that simultaneously minimizes the descriptive complexity of the tree(s) plus the data when encoded using those tree(s). In practice this is the hypothesis that can compress the data the most. We show not only that phylogenetic compression is an efficient method for encoding most phylogenetic data sets and is more efficient than compression schemes designed for single sequences, but also that it provides a clear information theoretic rule for determining when a collection of conflicting trees is a better explanation of the data than a single tree. By casting the parsimony problem in this more general framework, we also conclude that the so-called total-evidence tree--the tree constructed from all the data simultaneously--is not always the most economical explanation of the data.