Mean-Variance Portfolios

In 1952, a 25-year-old PhD student named Harry Markowitz published a twelve-page paper in the Journal of Finance and quietly changed the way the world invests. His insight was simple but powerful: you should not pick stocks one by one on the basis of expected return alone — you should pick combinations of assets, because the way their prices move together determines how much of their individual risk cancels out.

The tool Markowitz handed us is mean-variance analysis. Every portfolio can be described by just two numbers: its expected return $\mu$ (the weighted average of what each asset is expected to earn) and its variance $\sigma^2$ (how wildly the portfolio's value swings). Rational investors want high $\mu$ and low $\sigma^2$ — and the remarkable fact is that you can navigate this tradeoff geometrically.

The set of portfolios that deliver the highest return for every level of variance traces a curve in $(\sigma, \mu)$ space called the efficient frontier. Every portfolio below that curve is dominated — you could get the same variance with better return, or the same return with less variance. Every portfolio on it is, in Markowitz's sense, optimal.

Markowitz won the Nobel Memorial Prize in Economics in 1990 for this framework. The underlying mathematics — quadratic programming — is a solved problem: you can compute the efficient frontier exactly in polynomial time.

The chart below shows the efficient frontier for a two-asset portfolio. Move the slider to change the weight $w$ of Asset A (Asset B gets $1-w$). The dot traces how risk ($\sigma$) and return ($\mu$) change as you blend the two assets.

<!-- {{c_desc}} -->
<div class="controls">
  <label for="weightSlider">{{lbl_weight}} <span id="wVal">50%</span></label>
  <input type="range" id="weightSlider" min="0" max="100" value="50" step="1" aria-label="{{lbl_weight}}">
  <div class="asset-info">
    <span class="tag a">{{lbl_asset_a}}</span>
    <span class="tag b">{{lbl_asset_b}}</span>
  </div>
</div>
<canvas id="chart" width="420" height="280" aria-label="{{lbl_chart_aria}}"></canvas>
<div id="info" class="info-box"></div>

/* {{c_layout}} */
* { box-sizing: border-box; }
body { font-family: system-ui, sans-serif; color: #222; margin: 0; padding: 4px 8px; }
.controls { display: flex; flex-direction: column; gap: 6px; margin-bottom: 10px; }
label { font-size: .85rem; font-weight: 600; color: #444; }
input[type=range] { width: 100%; accent-color: #1d3557; cursor: pointer; }
.asset-info { display: flex; gap: 8px; }
.tag { font-size: .8rem; font-weight: 700; padding: 2px 8px; border-radius: 999px; }
.tag.a { background: #d0e8f7; color: #1a5276; }
.tag.b { background: #fde8d0; color: #784212; }
canvas { display: block; width: 100%; max-width: 420px; border-radius: 8px;
         background: #f7f9fb; border: 1px solid #dde4ec; }
.info-box { margin-top: 8px; font-size: .88rem; line-height: 1.55;
            background: #eef4fb; border-left: 3px solid #1d3557;
            padding: 8px 12px; border-radius: 4px; min-height: 3em; }

// Code not found

Notice how the curve bends inward — that leftward bulge is diversification at work. Because the two assets are not perfectly correlated ($\rho < 1$), mixing them reduces variance below a simple straight-line blend. The leftmost point of the frontier is the minimum-variance portfolio: the exact weight that minimizes risk. Any weight that lands on the upper half of the curve is efficient; the lower half is dominated.

The classical Markowitz problem — minimize $w^T \Sigma w$ subject to $w^T \mathbf{1} = 1$ and $w^T \mu = \mu_{\text{target}}$ — is a convex quadratic program. Convex quadratic programs are solvable in polynomial time; in practice, commercial solvers handle thousands of assets in milliseconds.

But real portfolios come with constraints that break convexity:

• Cardinality constraints ("hold at most $k$ assets") turn the problem into a mixed-integer program, which is NP-hard in general.
• Transaction costs (paying a fixed fee per asset traded) introduce binary variables, again NP-hard.
• Short-selling bans and lot sizes add integer and combinatorial structure.
• Robust optimization (the covariance matrix $\Sigma$ is uncertain) leads to semi-definite programs — still polynomial but much heavier.

In practice, portfolio managers use heuristics, relaxations, and branch-and-bound solvers rather than exact algorithms when constraints are heavy. The elegant polynomial core lives at the center; the messy NP-hard shell wraps real-world investing around it.

The statistical estimation problem is also hard in a different sense: estimating $\Sigma$ from historical data requires $O(n^2)$ parameters for $n$ assets, and with $n = 500$ assets and only a few years of data, the sample covariance matrix is nearly singular — leading to wildly unstable frontier estimates. Modern techniques like shrinkage estimators (Ledoit-Wolf, 2004) and factor models are needed to make the frontier numerically stable.

Markowitz's framework spread far beyond academia and now underpins much of the financial industry:

• Index funds and passive investing: the Capital Asset Pricing Model (CAPM), which extends mean-variance theory, justifies holding the entire market portfolio. John Bogle's Vanguard index funds — now managing trillions of dollars — rest on this theoretical foundation.
• Risk-parity strategies: funds like Bridgewater's All-Weather allocate by risk contribution rather than dollar weight, a direct generalization of the minimum-variance idea.
• Factor models: the Fama-French three-factor and five-factor models decompose expected returns into systematic risk factors, making the covariance matrix tractable for large universes of assets.
• Robo-advisors: platforms like Betterment and Wealthfront run mean-variance optimization continuously to keep millions of retail portfolios on the efficient frontier.
• Algorithmic trading: quantitative hedge funds solve portfolio optimization problems thousands of times per day, folding in transaction costs, market impact, and alpha signals.

The core idea — that risk is a quantity you can measure and trade against return — also shaped thinking far outside finance: in engineering reliability, drug combination therapies, and multi-objective optimization more broadly.

Harry Markowitz's mean-variance framework turned a vague intuition — "don't put all your eggs in one basket" — into a precise geometric object: the efficient frontier. The mathematics is a convex quadratic program, solvable in polynomial time, and the key message is that diversification is free lunch: by mixing assets with imperfect correlation, you reduce variance without sacrificing expected return.

The real world complicates the picture — cardinality constraints, uncertain covariance estimates, and transaction costs push the problem into NP-hard territory — but the elegant core remains. Every index fund, robo-advisor, and risk-parity strategy in existence owes something to that twelve-page paper from 1952.

The efficient frontier is a reminder that the best answer is not always extreme: sometimes the wise choice lives at a careful blend, not at any corner. That is a lesson as useful in optimization as it is in life.