There is no universally best optimizer
Portfolio optimization sounds like it should have one right answer: feed in the numbers, get the optimal weights. In reality there is a whole catalog of approaches, each built on different assumptions, optimizing for different objectives, and failing in different ways. The skill is not knowing one formula — it is choosing the right pattern for the situation, understanding its trade-offs, and comparing candidates fairly on the same portfolio before committing.
The classic and its discontents: mean-variance
Mean-variance optimization, the foundation of modern portfolio theory, finds the weights that maximize expected return for a given level of risk. It is elegant and Nobel-winning, and it has a well-known weakness: it is exquisitely sensitive to its inputs, especially expected returns. Tiny changes in return estimates produce wildly different "optimal" portfolios, often absurdly concentrated. Because expected returns are notoriously hard to forecast, naive mean-variance can amplify estimation error into confident-looking but fragile allocations. It is the right starting point and the wrong place to stop.
Approaches that lean on risk, not return forecasts
Much of the field exists to reduce dependence on unreliable return forecasts by leaning on risk structure, which is more estimable:
- Minimum variance ignores return forecasts entirely and simply seeks the lowest-risk portfolio. Useful when you trust your risk model far more than your return views.
- Risk parity allocates so that each asset contributes equal risk to the portfolio, rather than equal dollars. It avoids the trap of a "diversified" portfolio whose risk is dominated by one volatile sleeve.
- Risk budgeting generalizes this: you deliberately assign how much risk each asset or factor is allowed to contribute, expressing views through a risk budget rather than a return forecast.
These are not strictly better than mean-variance — they trade one set of assumptions for another. They are more robust to bad return estimates and, in exchange, give up the explicit return optimization that mean-variance provides.
Objectives, risk measures, and constraints
Every optimization pattern is defined by three choices worth understanding before running one:
- the objective — maximize return, minimize risk, maximize risk-adjusted return, or hit a risk budget;
- the risk measure — variance is standard, but downside-focused measures like CVaR penalize tail losses rather than symmetric volatility;
- the constraints — position limits, sector caps, turnover budgets, and tax-awareness that keep the "optimal" answer actually investable.
The same universe of assets produces very different portfolios depending on these three knobs, which is why understanding a pattern's defaults matters as much as running it.
The discipline of fair comparison
Because no approach is universally best, the most useful thing you can do is compare a few candidates on the same portfolio, under the same constraints — the same scope, the same data, the same fairness key. Run mean-variance, minimum-variance, and risk parity against your actual holdings and look at the resulting weights, the risk metrics, the diagnostics, and the suggested trades side by side. The comparison surfaces the trade-offs concretely: one approach concentrates, another diversifies risk, a third minimizes turnover. The "right" answer is the one whose trade-offs match your objective — and you can only see that by comparing.