Investors do not hold optimal portfolios. We use an experimental method to isolate factors that compel individuals to hold optimal portfolios. Our design includes two risky assets with perfectly negatively correlated payoffs so that all risk can be eliminated. We find that participants’ holdings approach optimal portfolios only under very specific conditions: the variance cost of holding an imbalanced portfolio is substantial and feedback on period-by-period outcomes is suppressed (eliminating the impact of cognitive biases resulting from misperceptions of randomness).