The Inputs Problem
Mean-variance optimisation is mathematically elegant but practically fragile. It requires precise estimates of expected returns, variances, and all pairwise covariances. For a portfolio of N assets, that is N expected returns, N variances, and N(N−1)/2 covariances — 1,275 distinct parameters for a 50-asset portfolio.
All of these must be estimated from historical data, and all of them are noisy. Expected return estimates from historical data are particularly unreliable: the standard error of a mean return estimated from 20 years of monthly data is large enough to make most estimates statistically indistinguishable from each other. The optimisation procedure is highly sensitive to these inputs: small changes in expected return assumptions can flip the "optimal" portfolio from heavily overweighting one asset to heavily underweighting it.
This fragility — the tendency of mean-variance optimisation to produce extreme, concentrated, and practically unusable portfolios — is one of the central criticisms of MPT in practice. It motivates approaches like the Black-Litterman model, which starts from the market portfolio as a neutral baseline and adjusts only in proportion to the investor's conviction in specific views.