Loanword formation seems to provide a good test bed for the growing field of computational phonology, since it occurs in a more tightly controlled environment than other language processing tasks. We show how feedforward neural networks and decision trees can be trained to predict the phonological structure of English loanwords in Japanese, and compare the performance of the two paradigms. In each case the system produces a phonemic representation of the Japanese form, after receiving as input the phonological feature matrix of the current and surrounding phonemes. The performance is improved with the inclusion of information about the stress pattern, orthography of reduced vowels and location of word boundaries.