Conventional investment wisdom advocates that investors should be well diversified, i.e., invest in as many assets as possible. We show instead that due to estimation errors in the inputs of portfolio strategies, the optimal level of diversification strikes a trade-off between accessing additional investment opportunities and limiting estimation risk, and thus is finite and can be relatively small. We also propose a set of procedures to select which assets are part of the restricted investment universe. Empirically, we show that limiting diversification with our method substantially outperforms unrestricted portfolios and makes portfolio theory valuable even in high-dimensional settings.