Resampled Frontiers vs Diffuse Bayes: An Experiment

Harry M. Markowitz and Nilufer Usmen

The experiment reported here compares two methods for handling uncertain inputs to a mean-variance analysis. Specifically, it compares Michaud’s resampled frontier versus Bayesian inference with diffuse prior. A simulated “referee” generates ten “truths” about 8 asset classes. For each truth it randomly generates one hundred histories. A simulated “Bayes Player” and “Michaud Player” process each history according to their respective methodologies, seeking portfolios to maximize given expected utility functions. Players are scored according to the actual utility achieved and their own estimates of this utility. The authors were surprised to find that, on average, the Michaud player won.