The core problem is that optimisation amplifies estimation errors, not just signal.
When you feed sample estimates of expected returns and covariances into a mean-variance optimiser, the maths treats those estimates as exact. But they aren't — every estimate has error. The optimiser systematically overweights assets whose Expected Return — definition">expected return estimates happen to be too high, and underweights assets whose estimates happen to be too low. [1]
The result: the output looks precise but is built on noise. Portfolios that appear optimal in-sample often concentrate heavily in whatever looked good during the estimation window — and subsequently underperform.
DeMiguel, Garlappi, and Uppal (2009) tested 14 optimisation strategies — including Markowitz, shrinkage estimators, and Bayesian approaches — against a simple 1/N equal-weight portfolio across 7 real datasets. The naïve 1/N portfolio was hard to beat. [1]
The punchline: to recover the diversification benefit of Markowitz optimisation, you would need 500 years of monthly data — far more than any investor has. [1]
| What optimisation gets wrong | What it gets right |
|---|---|
| Precise weights (39.7% vs. 40%) | Low-correlation assets genuinely reduce risk |
| Concentrated tilts toward recent winners | The diversification principle itself |
| False sense of precision | Direction of allocation (rough bands) |
Three takeaways: [1]
Even before optimisation runs, your Capital Market Assumptions (CMAs) drive the output. If you assume equity earns 15% instead of 11%, the model pours money into equity. Small errors in expected return inputs produce large swings in portfolio weights. [5]
Apply this → Go to Portfolio Builder to model your own allocation using rough bands and sensible CMAs — not optimiser-generated precision weights.