Mean-Variance Portfolio

What Markowitz actually did in 1952

Before March 1952, "diversification" was folk wisdom. Investors split their money across stocks because their grandparents told them to, but no one had written down what diversification was for or how much of it was enough. Harry Markowitz, then a 25-year-old PhD student at the University of Chicago, did something almost shockingly mundane: he treated portfolio choice as a constrained optimisation problem. The result, published as "Portfolio Selection" in the Journal of Finance in March 1952, ran fourteen pages and contained no novel mathematics. What it had was a new object: the portfolio as a vector of weights, evaluated on two numbers — its expected return and its variance — chosen by trading one off against the other.

That single move made finance a quantitative discipline. The 1952 paper led directly to Sharpe's 1964 CAPM, the entire index-fund industry, the Sharpe ratio, Black-Litterman, risk parity, factor investing, and the trillions of dollars now managed against benchmarks that exist because mean-variance gave us the language of a "benchmark" in the first place. Markowitz shared the 1990 Nobel Memorial Prize in Economic Sciences with William Sharpe and Merton Miller. The committee's citation called it "the theory of choice under uncertainty." It is, more practically, the theory that says covariance is what matters.

The formal problem

Let there be N risky assets with expected return vector μ ∈ ℝ^N and covariance matrix Σ ∈ ℝ^(N×N) (symmetric positive definite). A portfolio is a weight vector w ∈ ℝ^N. Its expected return and variance are

μ_p = w'μ
σ_p² = w'Σw

The mean-variance optimisation problem is

minimise   w'Σw
subject to w'μ = μ*       (target return)
           w'1 = 1        (weights sum to one)

Form the Lagrangian L = w'Σw − 2λ₁(w'μ − μ*) − 2λ₂(w'1 − 1) and set ∂L/∂w = 0. The first-order condition is Σw = λ₁μ + λ₂1, which solves to

w*(μ*) = Σ⁻¹ (λ₁ μ + λ₂ 1)
       = a + b · μ*

where a and b are constant vectors built from Σ⁻¹, μ, and the ones vector through three scalars A = 1'Σ⁻¹μ, B = μ'Σ⁻¹μ, C = 1'Σ⁻¹1, D = BC − A². Two key facts fall out:

• Two-fund theorem. Every efficient portfolio is a linear combination of just two reference portfolios — say the minimum-variance portfolio and any other frontier portfolio. You can build the entire frontier by mixing those two.
• The frontier is a parabola. Substituting w*(μ*) back into the variance gives σ_p² = (C μ*² − 2A μ* + B) / D, a parabola in (σ_p², μ_p) space. In (σ_p, μ_p) space it is the right branch of a hyperbola; the upper half is the efficient frontier.

Why diversification works — the role of ρ

Take two assets. Their portfolio variance with weights w and (1−w) is

σ_p² = w² σ_1² + (1−w)² σ_2² + 2 w (1−w) ρ σ_1 σ_2

where ρ is their correlation. The cross-term is the entire show. If ρ = 1 the portfolio's standard deviation is just the weighted average of the individual standard deviations — there is no diversification gain. If ρ = 0 the variance is reduced by mixing because the covariance term vanishes. If ρ < 0 the variance can be driven below either asset's individual variance — the textbook "two negatively correlated stocks can be combined into a less-risky portfolio than either alone" result.

In the (σ, μ) plane, as you slide w from 0 to 1 with ρ < 1, the curve bulges left relative to the straight line you'd draw at ρ = 1. That bulge is diversification visualised. Lower ρ means a bigger bulge; ρ = −1 lets the curve pinch all the way down to σ = 0 at a specific mix. The efficient frontier is what you get when this story is told for all N assets at once.

The riskless asset and the tangency portfolio

Now add a riskless asset with return r_f and zero variance. Any mix of cash and a risky portfolio p has return r_f + α(μ_p − r_f) and standard deviation α σ_p — a straight line in (σ, μ) space starting at (0, r_f) and passing through (σ_p, μ_p). The line you want is the steepest one that still intersects the feasible set of risky portfolios. That is the line tangent to the efficient frontier; its tangent point is the tangency portfolio.

The slope of this line is the Sharpe ratio (μ_p − r_f)/σ_p of the tangency portfolio — by construction the maximum Sharpe ratio attainable from risky assets. The line itself is called the Capital Market Line, and it dominates every interior point of the original frontier: for any portfolio strictly inside the frontier, there is a cash-plus-tangency mix with the same volatility and a higher expected return.

From this comes Tobin's separation theorem (James Tobin, 1958): every mean-variance investor with access to the riskless asset holds the same portfolio of risky assets — the tangency portfolio — and only varies the cash-vs-risky split based on personal risk tolerance. The composition of risky holdings is separated from the appetite for risk. CAPM extends this with one further step: in equilibrium, the tangency portfolio is the value-weighted market portfolio.

Closed-form solution with a riskless asset

With a riskless asset, the constraint w'1 = 1 disappears (cash absorbs whatever is left), and the problem reduces to maximising the Sharpe ratio over risky weights. The closed-form solution is

w_tangency ∝ Σ⁻¹ (μ − r_f · 1)

This is the single most-cited equation in quantitative portfolio management. Note what it does and does not depend on:

• It depends on Σ⁻¹ (μ − r_f·1). Both the inverse covariance and the excess-return vector enter. The proportionality constant is set by the chosen total risk; usually you normalise so weights sum to one.
• Small changes in μ can produce huge swings in w. Σ⁻¹ is a noise amplifier. Two assets with nearly identical estimated returns and high estimated correlation will receive enormous opposite weights from the unconstrained optimum — this is the source of the "garbage in, garbage out" reputation.

Worked example: two assets

Suppose equities have μ_e = 8%, σ_e = 18%; bonds have μ_b = 3%, σ_b = 6%; their correlation is ρ = 0.1 and the risk-free rate is 2%. The covariance matrix is

Σ = [ 0.0324   0.00108 ]
    [ 0.00108  0.0036  ]

Σ⁻¹ ≈ [ 31.4   −9.4  ]
      [ −9.4   282.7 ]

μ − r_f = [ 0.06 ]
          [ 0.01 ]

Σ⁻¹ (μ − r_f) ≈ [ 1.79 ]
                 [ 2.27 ]

w_tangency ∝ (1.79, 2.27)  →  normalise:
w_tangency ≈ (0.44, 0.56)

So a mean-variance investor with these inputs holds 44% in stocks and 56% in bonds, then levers or unlevers that mix with cash to hit their personal risk tolerance. The tangency portfolio's expected excess return is 0.44·0.06 + 0.56·0.01 = 3.20% over the risk-free rate; its standard deviation is √(w'Σw) ≈ 8.4%; its Sharpe ratio is 3.20/8.4 ≈ 0.38. Bumping ρ from 0.1 to 0.5 leaves the math the same shape but moves the tangency weights to roughly (0.55, 0.45) — a non-trivial sensitivity to a covariance input that is itself only loosely estimated.

Estimation error — the biggest practical problem

The mean-variance program assumes μ and Σ are known. They are estimated, and they are estimated badly. The standard error on an annual mean from T years of data is roughly σ/√T. For a typical equity with σ ≈ 20% and T = 25 years, that is 4% — about the same magnitude as the equity premium itself. You cannot distinguish a "good" stock from a mediocre one from a 25-year track record. Σ is more accurately estimated from high-frequency data but is non-stationary: correlations regularly jump from 0.2 to 0.9 in crises.

Two empirical consequences:

• Extreme concentration. Unconstrained Markowitz portfolios load 200% long on one or two assets and 100% short on others. Add a no-shorting constraint and the optimum often collapses to one or two long positions. The Σ⁻¹ in the closed-form is mathematically the culprit: small noise in eigenvalues near zero turns into enormous weights.
• Out-of-sample failure. DeMiguel, Garlappi and Uppal ("Optimal Versus Naïve Diversification: How Inefficient Is the 1/N Strategy?", Review of Financial Studies, 2009) showed across seven datasets that classical Markowitz portfolios underperform the equal-weight 1/N rule out-of-sample on Sharpe ratio, certainty-equivalent return, and turnover. Sample-mean-driven optimisation, in their data, was strictly dominated by a rule a five-year-old could follow.

Modern fixes

Method	Year	What it does	Limitation
Sample covariance + shrinkage	Ledoit-Wolf 2003	Average sample Σ with a structured target (e.g. constant correlation); shrinkage chosen to minimise MSE	Still uses noisy sample μ
Robust optimisation	Goldfarb-Iyengar 2003	Optimise worst-case over an uncertainty set on μ and Σ	Conservative; requires uncertainty-set tuning
Bayesian / Black-Litterman	Black-Litterman 1990	Start from market-implied μ; blend in views with confidence weights	View specification is subjective
Risk parity	Qian 2005; Bridgewater 1996	Allocate so each asset contributes equally to risk; ignore μ entirely	Implicitly assumes equal Sharpes; tends bond-heavy
Hierarchical risk parity	Lopez de Prado 2016	Use clustering on Σ to allocate without inverting it	Heuristic; weaker theoretical foundation
Resampled efficiency	Michaud 1998	Bootstrap many (μ, Σ) draws; average the optimal weights	Patented; computationally heavy
Factor-based optimisation	Fama-French 1992+	Replace asset-level Σ with a low-dimensional factor model	Model risk in the factor structure
Minimum-variance portfolio	—	Drop the μ constraint; just minimise w'Σw	Throws away all information about expected returns

The dominant institutional practice today is a hybrid: estimate Σ with Ledoit-Wolf shrinkage or a factor model, anchor μ on Black-Litterman with market priors and a small number of views, optimise with realistic constraints (no leverage, position limits, turnover penalties), and pressure-test the resulting portfolio under historical stress scenarios. The pure 1952 problem is rarely solved as stated; what survived was the framework — that decisions should be made on the joint distribution of returns, not on individual assets in isolation.

Why "mean-variance" is not "everything that matters"

The framework reduces an investor's preference to a function of just two moments. That is exactly correct in two cases: when returns are jointly Gaussian, or when utility is quadratic. Neither is empirically right.

• Fat tails. Stock returns have kurtosis around 4-15 versus 3 for Gaussian. The 1987 crash was a 22-σ event under a Gaussian-fit model; under any honest empirical model it was a routine fat-tail draw. Mean-variance underweights this risk because variance treats large losses and large gains symmetrically.
• Skewness. Most asset returns are negatively skewed (small frequent gains, occasional large losses). Investors plausibly care about skew, but mean-variance does not see it.
• Correlation breakdown. ρ is not a constant. In the 2008 crisis essentially every risky asset class correlation went to one, eviscerating diversification just when investors most needed it.
• Time-varying μ. Predictability through valuation ratios (Campbell-Shiller), momentum (Jegadeesh-Titman), and macro variables means μ depends on the current state of the world — but the static 1952 problem treats it as a constant.

These limitations motivated downside-risk measures (semivariance, conditional value-at-risk), higher-moment optimisation (mean-variance-skewness-kurtosis), and the modern factor-based approach where Σ is decomposed into low-dimensional factor exposures plus residual.

Legacy: what mean-variance gave us

• The Sharpe ratio. (μ − r_f)/σ. The number-one performance metric in finance, defined as the slope of the line from r_f to a portfolio in (σ, μ) space — exactly the object mean-variance maximises.
• CAPM. Sharpe (1964) added equilibrium to Tobin's separation: if every investor holds the tangency portfolio, in equilibrium the tangency must be the market itself, which forces E[R_i] = r_f + β_i(E[R_m] − r_f). The line in σ-μ space becomes the security market line in β-μ space.
• Index funds. Vanguard's 1976 First Index Investment Trust was a direct application of the logic: if the market portfolio is the tangency, owning the market is mean-variance optimal.
• Risk-budgeting. Bridgewater All Weather (1996), AQR Style Premia, every modern "risk parity" product is a particular response to mean-variance's estimation problem.
• Modern Portfolio Theory (MPT). The umbrella term used in textbooks and CFA curricula to cover everything that descends from the 1952 paper.

Common pitfalls

• Optimising on historical means. The single most common error. Use of the trailing 10-year mean as μ has produced spectacularly bad portfolios at almost every major market turn. Industry practice — for good reason — uses market-implied or model-implied μ instead.
• Treating Σ as fixed. Covariances move. A portfolio optimised on a low-volatility regime can become wildly off-target when the regime shifts. Rolling-window or DCC-GARCH estimates partially address this.
• Ignoring transaction costs. A vanilla MV optimum can have 200% turnover per year. Adding a turnover or rebalancing-cost penalty changes the problem substantially.
• Inverting near-singular Σ. When N is large relative to T, the sample Σ has zero or near-zero eigenvalues; Σ⁻¹ is undefined or astronomical. Shrinkage, factor models or pseudo-inverses are required.
• Confusing the efficient frontier with the investable frontier. The unconstrained frontier allows arbitrary shorts and leverage. Real-world frontiers under no-shorting, leverage caps, and sector limits look very different — and the tangency portfolio under constraints can shift dramatically.