Markowitz Portfolio Theory: The Efficient Frontier

In 1952, Harry Markowitz published a 14-page paper that revolutionized finance. His insight: investors care about two things — expected return and risk (variance). Combining assets with imperfect correlation reduces total risk without sacrificing return. The efficient frontier is the set of optimal portfolios that maximize return for each level of risk.

1. Return and Risk

For a single asset, we model:

Expected return: μ = E[r] = (1/T) Σ r_t (historical mean) Variance: σ² = E[(r - μ)²] (measure of dispersion) Std deviation: σ = \sqrtσ² (annualized: σ_daily \times \sqrt252)

Historical annual standard deviations (rough): US Treasury bills ~1%, bond index ~7%, S&P 500 ~15–18%, individual tech stocks ~30–50%, Bitcoin ~70–90%.

The assumption that variance captures all relevant risk is a simplification — returns are not normally distributed (fat tails, skewness) and correlation between assets changes in crises.

2. Covariance and Correlation

The covariance between assets i and j measures how returns move together:

Cov(r_i, r_j) = E[(r_i - μ_i)(r_j - μ_j)] Correlation: ρ_ij = Cov(r_i, r_j) / (σ_i \cdot σ_j) \in [-1, 1] ρ = +1: perfectly correlated \to no diversification benefit ρ = 0: uncorrelated \to partial diversification ρ = -1: perfectly negatively correlated \to maximum diversification

For N assets, we need N expected returns, N variances, and N(N−1)/2 covariances. For 500 S&P stocks: 124,750 covariances. The full covariance matrix is N×N, estimated from historical data (and notoriously unstable).

3. Portfolio Math

A portfolio is defined by weights w = (w₁, w₂, ..., wₙ) where Σwᵢ = 1:

Portfolio expected return: μ_p = wᵀ \cdot μ = Σ wᵢ μᵢ Portfolio variance: σ²_p = wᵀ Σ w = Σᵢ Σⱼ wᵢ wⱼ Cov(rᵢ, rⱼ) For 2 assets (weights w, 1-w): σ²_p = w²σ₁² + (1-w)²σ₂² + 2w(1-w)ρσ₁σ₂

The key insight: when ρ < 1, σ_p < w·σ₁ + (1−w)·σ₂. The portfolio's risk is less than the weighted sum of individual risks. This is the free lunch of diversification — reducing risk without giving up expected return (provided assets are not perfectly correlated).

Two-asset example: TSLA (μ=30%, σ=55%) and TLT bonds (μ=5%, σ=15%). With ρ = −0.3, a 60/40 TSLA/TLT portfolio has σ_p ≈ 33% — lower than 60%×55% + 40%×15% = 39%.

4. The Efficient Frontier

For every possible target return μ_p, solve the quadratic program:

Minimize: wᵀ Σ w (portfolio variance) Subject to: wᵀ μ = μ_p (target return) Σ wᵢ = 1 (weights sum to 1) [optionally: wᵢ \geq 0 for long-only]

Sweeping μ_p from minimum to maximum traces out the minimum variance frontier — the parabola in (σ_p, μ_p) space. The upper half above the global minimum variance portfolio is the efficient frontier: for each risk level, these portfolios offer the maximum possible return.

Any rational, risk-averse investor should hold a portfolio on the efficient frontier. Which point depends on individual risk tolerance — more risk-tolerant investors move up the frontier (higher return, higher risk).

5. Sharpe Ratio and the Market Portfolio

The Sharpe ratio measures return per unit of risk taken:

Sharpe = (μ_p - r_f) / σ_p r_f = risk-free rate (3-month Treasury bill yield) Typical targets: S&P 500 \approx 0.5-0.8 historically Warren Buffett Berkshire ~0.6-0.8 over 50 years

The portfolio with the maximum Sharpe ratio is the tangency portfolio — the point on the efficient frontier where a line from the risk-free rate just touches the frontier. This is the market portfolio under CAPM assumptions.

By the Mutual Fund Separation Theorem: any investor can achieve the optimal risk-return tradeoff by combining the tangency portfolio (e.g., a total market index fund) with the risk-free asset. More risk tolerance → more equity, less bonds — but always the same equity portfolio.

6. CAPM

The Capital Asset Pricing Model (Sharpe 1964, Lintner 1965) extends Markowitz to an equilibrium model. If all investors hold mean-variance efficient portfolios, the market must clear — the aggregate portfolio held by all investors is the market portfolio.

E[r_i] = r_f + β_i \cdot (E[r_M] - r_f) β_i = Cov(r_i, r_M) / σ²_M (systematic risk) Market risk premium E[r_M] - r_f \approx 5-7% historical (US equity) Example: stock with β=1.5 should return r_f + 1.5 \times (market premium) At r_f=5%, premium=6%: expected return = 5% + 9% = 14%

Beta measures how much market risk an asset contributes. Only systematic (market-correlated) risk is rewarded with higher expected return — idiosyncratic risk can be diversified away for free, so investors won't pay a premium for bearing it.

7. Criticisms and Extensions

Sensitivity to inputs: Small changes in expected return estimates produce large changes in optimal weights. The optimizer is "an error maximizer" — input noise amplifies. Solutions: Black-Litterman model (shrinkage), robust optimization.
Normal distribution assumption: Real returns have fat tails (kurtosis > 3) and crash correlations rise (stocks correlate more during market stress). Markowitz assumed stationary covariance.
Expected return estimation: Predicting future returns from historical data is notoriously unreliable. "Garbage in, garbage out."
Factor models (Fama-French): Three-factor model adds size (SMB) and value (HML) factors to beta. Five-factor adds profitability and investment. Better description of returns cross-section.
Risk parity: Allocate by equal risk contribution rather than by weight. Each asset contributes equally to portfolio variance. Often more robust than Markowitz.

Despite its assumptions, Markowitz's framework remains the foundation of institutional asset management. The S&P 500's dominance of passive investing implicitly relies on CAPM's market portfolio insight.