Mean-Variance Optimisation Explained: Markowitz, the Efficient Frontier, and Why Real Portfolios Use Heuristics

In 1952, Harry Markowitz published "Portfolio Selection" in the Journal of Finance, a paper that transformed investment management from an art practised through intuition into a discipline with mathematical foundations. His insight — that the risk of a portfolio is not the weighted average of individual asset risks, but depends critically on how those assets move relative to each other — is so fundamental that it is easy to forget it was once a novel idea.

Mean-variance optimisation (MVO), the framework derived from Markowitz's insight, remains the theoretical foundation of modern portfolio construction. It also has well-documented practical limitations that explain why the most sophisticated real-world portfolios are built using a combination of quantitative tools and experienced judgement rather than raw optimisation. This guide explains the theory, the limitations, and the practical alternatives.

The Core Insight: Diversification as a Free Lunch

Before Markowitz, investment practitioners thought about portfolio construction by focusing on the risk and return of individual securities. Good diversification meant simply holding many different assets.

Markowitz demonstrated that the relevant risk of a new asset, when added to a portfolio, is not its own standalone volatility but its correlation with the existing portfolio. A highly volatile asset that is uncorrelated with (or negatively correlated with) the existing portfolio can actually reduce portfolio risk when added, even though it appears risky in isolation. This is the mathematical basis for the oft-cited observation that diversification is "the only free lunch in investing."

The practical implication is that the relevant question when evaluating any asset is not "how risky is this?" but "how does this change the risk of my overall portfolio?"

The Inputs to MVO

Mean-variance optimisation requires three sets of inputs:

Expected returns: The expected annualised return for each asset class under consideration. This is the most difficult input to estimate. Forecasting returns is notoriously imprecise; even sophisticated multi-factor models have large uncertainty bands around their predictions.

Expected volatilities: The expected standard deviation of returns for each asset class. Historical volatility is the most common input, with adjustments for structural changes. This input is more stable than expected returns over time.

Correlations: The expected pairwise correlations between all asset classes. Historical correlations are again the usual input, though correlations are not stable and tend to rise in crisis conditions (a painful property for investors seeking diversification precisely when it is most needed).

Given these inputs, MVO calculates the set of portfolios that, for any given level of expected return, minimise expected portfolio variance — or equivalently, for any given level of variance, maximise expected return. This set of portfolios is called the efficient frontier.

The Efficient Frontier

The efficient frontier is typically depicted as a curve in return-volatility space, with the y-axis representing expected return and the x-axis representing expected volatility (standard deviation). Every portfolio on the frontier is "efficient" in the sense that no other portfolio exists with the same expected return and lower expected volatility, or the same expected volatility and higher expected return.

Portfolios below the frontier are dominated — they can be improved by choosing a frontier portfolio with the same risk but higher return, or the same return but lower risk. Portfolios above the frontier are impossible to construct given the available assets and the stated inputs.

The efficient frontier is derived for a specific set of assets: different asset universes produce different frontiers. A frontier computed using only equities is very different from one that includes bonds, real assets, and alternatives.

Why MVO Portfolios Often Look Extreme

In theory, MVO produces beautifully optimal portfolios. In practice, the portfolios produced by unconstrained optimisation are often highly concentrated in a small number of assets, with extreme weights that would be immediately rejected by any experienced portfolio manager.

The root cause is a property sometimes called "error maximisation". MVO finds portfolios that appear optimal given the stated inputs. But the inputs themselves — expected returns, volatilities, correlations — are estimated with substantial uncertainty. Small changes in these estimates can produce dramatically different optimal portfolios. The optimiser, blind to input uncertainty, confidently weights the portfolio heavily towards the assets with the most attractive apparent estimates — precisely the assets for which estimation error is most likely to be largest.

A famous illustration: MVO applied to historical asset class returns in any historical sample tends to recommend very high allocations to the asset classes that happened to perform well in that particular sample and very low allocations to those that performed poorly. This is an artefact of fitting to past data, not a reliable guide to future expected returns.

The Black-Litterman Model

Fischer Black and Robert Litterman at Goldman Sachs developed an important improvement to raw MVO in 1990. The Black-Litterman model addresses the estimation error problem by starting from an equilibrium set of expected returns — typically those implied by the current market capitalisation weights (what the market collectively "expects" given current prices) — and adjusting from this neutral starting point only according to the investor's explicit views about specific markets.

The result is portfolios that are more diversified and stable than raw MVO output, because the default position (market-weight portfolio) is already well-diversified and each departure from it requires an explicit view with a stated confidence level.

Black-Litterman is widely used by institutional asset managers. Its advantage is that it forces explicit articulation of views and their confidence levels, rather than allowing estimation noise in expected return inputs to drive extreme allocations.

Constraint-Based Optimisation

A simpler and more widespread practical approach is to apply constraints to the optimisation that prevent it from producing extreme portfolios. Common constraints include:

• Maximum and minimum weights for each asset class (e.g., no single equity market exceeds 30% of the portfolio; minimum 5% in each included asset class).
• Tracking error constraints relative to a benchmark.
• Turnover constraints that limit the degree of change between rebalancing periods.
• Factor exposure constraints (e.g., portfolio beta between 0.8 and 1.2 relative to the market).

Constrained MVO produces portfolios that are closer to the unconstrained optimum than pure heuristics but that are not distorted by estimation error. Most institutional portfolio management systems implement some form of constrained optimisation.

Liability-Driven Investing: An Alternative Framework

For investors with explicit future liabilities — pension funds with defined benefit obligations, endowments with spending commitments, insurance companies with policy payments — mean-variance optimisation relative to total return is only one possible objective. Liability-driven investing (LDI) reformulates the portfolio construction problem as matching or hedging the present value of liabilities rather than maximising absolute return.

In an LDI framework, the optimal portfolio is not the one that maximises the Sharpe ratio but the one that minimises the volatility of the surplus (assets minus liabilities). This leads to very different portfolio structures: UK defined benefit pension schemes following LDI approaches typically hold large allocations to long-duration gilts, which match the interest rate sensitivity of their long-dated pension obligations, combined with a growth portfolio designed to generate excess return above the liability hurdle.

For individual investors, the LDI framework is less directly applicable but the underlying insight — that risk should be measured relative to future obligations rather than in absolute terms — is valuable. An investor planning to fund school fees in GBP in five years has a different "risk-free" asset than one planning to retire in Australia in 20 years.

Why Most Real Portfolios Use Heuristics

Given the limitations of MVO, most real-world portfolio construction — from high-street financial planning to sophisticated private banking — uses a combination of:

Rule-of-thumb asset allocations: 60/40, risk-based guidelines calibrated to client age and risk tolerance, target-date glidepaths. These are not mathematically optimal but are stable, understandable, and robust to the parameter uncertainty that undermines MVO.

Strategic asset allocation (SAA) frameworks: Long-run target allocations based on expected risk premia, implemented with rebalancing bands rather than continuous optimisation. This captures the broad insight from MVO (diversification is valuable) without requiring precise parameter estimates.

Factor-based construction: Identifying desired factor exposures (market beta, value, quality, duration, liquidity premium) and constructing portfolios that achieve those exposures at minimum cost and maximum diversification within each factor.

Judgement and experience: The inputs to any portfolio construction process include qualitative judgements about structural changes in markets, regulatory environment, and macroeconomic conditions that are not reliably captured by historical statistics.

Sophisticated institutions use quantitative optimisation as one tool within a broader construction process, not as the sole determinant of portfolio weights.

Compliance and Regulatory Note

Portfolio optimisation models depend on inputs (expected returns, volatilities, correlations) that are inherently uncertain. Efficient frontier portfolios constructed using historical data are not guaranteed to be optimal in future periods. All investment involves risk, including the risk of capital loss. Past performance is not a reliable indicator of future results. This article is for information only and does not constitute personal financial advice.

How Global Investments Can Help

Portfolio construction for high-net-worth international investors requires integrating mathematical rigour with practical experience and judgement. At Global Investments, we draw on both quantitative tools and decades of experience across market cycles to construct portfolios that are genuinely diversified, appropriately allocated to client circumstances, and robust to the uncertainty inherent in any forward-looking investment process. If you would like to discuss how your current portfolio allocation was constructed and whether it reflects an evidence-based approach to risk and diversification, please contact our team.