Two issues are discussed in this paper. The first is whether a formal definition and justification of simplicity (parsimony) in scientific inference can be found, and whether an optimal level of simplicity is obtainable. A definition of simplicity is possible, as are the optimum conditions for the desired degree of simplicity. The model of inference used here relates Bayesian inference to algorithmic information theory. Simplicity is examined in the light of induction, the Duhem-Quine thesis, and bounded rationality. The second issue relates to the role that simplicity might play in econometric modeling. This is elucidated with some remarks on the 'general to specific' approach to modeling and discussions on the purpose of a model. Copyright 1995 by Royal Economic Society.