Abstract In this paper we compare practical procedures for implementing the well-known “growth-optimal” strategy for managing portfolios. The policy involves choosing in each investment period the portfolio that maximizes the expectation of the logarithm of the portfolio's return. We consider two vesions of the decision rule—the expected-log function itself and two-monent approximation of it—and three procedures for determining ex ante expectations of returns. In addition to a nonparametric approach, we examine two parametric models, one that takes returns to be jointly lognormal and a second that imposes a single-factor structure on the logs of returns. The relative performance of the six procedures is compared using data for indexes of securities and for the individual stocks in the Dow-Jones Industrial Average. Our main conclusion is that relatively simple methods for carrying out the growth-optimal program work about as well as methods that are computationally difficult. While naively analyzing historical returns cannot be relied upon to provide above-average performance, there are indications that performance can be improved by giving higher weight to the most recent returns.