for formulating views, and that the optimal portfolio is simply one that tilts with some set of weights on the view portfolios, it is probably easier to specify weights on those tilt portfolios directly rather than to specify expected returns, degrees of confidence, and correlations between views. There are, however, at least two reasons why the Black-Litterman model is necessary. First, if one simply specifies weights on view portfolios, one loses the insights that Black-Litterman brings concerning the effects of the different parameters on the optimal weights. Of course that loss has to be balanced against the difficulty of knowing how to set those parameters in the first place. Since the original Black -Litterman paper was written, I have often received the question, "How do you determine the omega matrix?" There is no simple or universal answer. We know what these parameters represent-the expected excess returns on the view portfolios, the degree of uncertainty in the views, and the correlations between views- but the right way to specify such information is certainly context dependent. When the views are the product of quantitative modeling, for example, the expected returns might be a function of historical performance, the degree of confidence might be set proportional to the amount of data supporting the view, and correlations between views might be assumed to be equal to the historical correlations between view portfolio returns. Other direct approaches to specifying weights on view portfolios can generally be mapped into particular assumptions on the expected excess returns and the omega matrix of Black-Litterman. At least in the context of Black-Litterman, the portfolio manager knows what these parameters represent, and can thus address the issue of whether those specifications make sense. The second, and perhaps more important, reason that the Black-Litterman framework really is necessary is because in the real world one hardly ever optimizes in an unconstrained environment. The real power of the Black-Litterman model arises when there is a benchmark, a risk or beta target, or other constraints, or when transaction costs are taken into account. In these more complex contexts, the optimal weights are no longer obvious or intuitive. The optimal portfolio is certainly not simply a set of tilts on view portfolios. Nonetheless, the manager can be confident that when the optimizer goes to work using the Black-Litterman expected excess returns, the same trade-off of risk and return-which leads to intuitive results that match the manager's intended views in the unconstrained case-remains operative when there are constraints or other considerations. Having made this point, it is nonetheless worth noting that, as shown in He and Litterman (1999), in a few special cases the optimal portfolios given constraints retain some intuitive properties. In our paper we consider in turn the case of a risk constraint, a leverage constraint, and a market exposure constraint. In the case of optimizing relative to a specified level of risk, the optimal portfolio is just a linearly scaled version of the solution of the unconstrained optimization problem. However, because of the scaling, the view portfolio deviations no longer tilt away from the market portfolio, but rather from a scaled market portfolio. Otherwise the intuition of the unconstrained portfolio remains. In the case of a fully invested, no-leverage constraint, a constraint where the portfolio weights sum to 1, another portfolio enters the picture. There exists a "global minimum-variance portfolio" that minimizes the risk of all portfolios that