Modern Portfolio Theory and the Efficient Frontier

Diversification is often called the only free lunch in finance. The slogan is misleading if you read it as "own many things." The free lunch is owning things that do not move together. Harry Markowitz turned that observation into a picture: the efficient frontier.

The picture is simple enough to draw on a napkin. The consequences are not. It tells you what risk and return are achievable, why correlation matters more than the number of holdings, and why a portfolio can be more than the sum of its parts. It also explains why the whole exercise is fragile: the frontier is drawn with estimates, and estimates are wrong.

Two Assets, One Formula

Take two assets:

• A: expected return 8%, volatility 16%.
• B: expected return 4%, volatility 6%.
• Correlation $\rho = 0.2$.

Put a fraction $w$ in A and $1-w$ in B. The portfolio's expected return is the weighted average:

$\mu_p = w\mu_A + (1-w)\mu_B.$

The volatility is not the weighted average, because the assets do not move in lockstep:

$\sigma_p^2 = w^2\sigma_A^2 + (1-w)^2\sigma_B^2 + 2w(1-w)\rho\,\sigma_A\sigma_B.$

The last term is the whole story. If $\rho = 1$, the assets move together and the formula collapses to the weighted average. If $\rho < 1$, the portfolio's risk is lower than the weighted average. If $\rho = -1$, the assets can cancel each other.

At a 50/50 split, the portfolio has a 6.00% expected return. Its volatility is 9.09%, while the weighted average of the two volatilities is 11.00%. Diversification removed 1.91 percentage points of risk without sacrificing any expected return. That is the free lunch.

The correlation dial controls how much is on the menu:

At $\rho = -1$, the two assets can be combined into a zero-risk portfolio: 27.3% in A and 72.7% in B. That portfolio returns about 5.09%, which is above the 2% risk-free rate. A riskless 5% return is an arbitrage; in real markets, correlations are never that cooperative.

The Minimum-Variance Portfolio

The 50/50 split is arbitrary. There is a weight that minimizes risk:

$w^* = \frac{\sigma_B^2 - \rho\sigma_A\sigma_B}{\sigma_A^2 + \sigma_B^2 - 2\rho\sigma_A\sigma_B}.$

For our example, that is 6.6% in A and 93.4% in B, with a volatility of 5.91% and a return of 4.26%. It is the leftmost point of the feasible set.

With long-only weights, the formula is clamped to $[0,1]$; if the unconstrained minimum is outside that range, the minimum is at a corner. At $\rho = 1$ and different volatilities, for example, the zero-risk portfolio requires shorting the more volatile asset; long-only, the minimum is simply the lower-volatility asset.

The minimum-variance portfolio is not the best portfolio. It is the safest risky portfolio. Everything to its right has higher risk; the useful part of the curve is the part above it.

The Efficient Frontier

Plot every possible portfolio in risk–return space. The result is a curve. The upper edge of the feasible set is the efficient frontier: for each level of risk, it gives the highest achievable expected return.

For two assets the feasible set is the curve itself. For many assets it is a filled region, and the frontier is its upper-left boundary. In either case:

• Portfolios below the frontier are dominated: you could get more return for the same risk, or less risk for the same return.
• The minimum-variance portfolio is the leftmost point.
• The efficient branch runs from the minimum-variance point up to the highest-return asset (or beyond, if leverage is allowed).
• The lower branch is inefficient: it has lower return and higher risk than the minimum-variance portfolio.

With $N$ assets and weights $w$, the same idea is written in matrix form:

$\mu_p = w^\top \mu, \qquad \sigma_p^2 = w^\top \Sigma w,$

where $\Sigma$ is the covariance matrix. The frontier is found by minimizing $\sigma_p^2$ for each target $\mu_p$ (or maximizing $\mu_p$ for each target risk), subject to the budget constraint $\mathbf{1}^\top w = 1$ and, usually, long-only constraints $w \ge 0$.

The frontier is a set of portfolios, not a single answer. Which point you want depends on how much risk you can tolerate.

Cash, Sharpe, and the Tangency

Now add a risk-free asset paying 2%. You can mix cash with any risky portfolio. The best mix is the one with the highest Sharpe ratio:

$S = \frac{\mu_p - r_f}{\sigma_p}.$

Geometrically, draw a line from the risk-free point $(0, r_f)$ to the risky frontier. The steepest such line touches the frontier at the tangency portfolio. For our example, that portfolio is 30.9% A and 69.1% B, with a return of 5.24%, a volatility of 7.06%, and a Sharpe ratio of 0.458.

The tangency portfolio is the maximum-Sharpe portfolio. Every efficient portfolio with cash is a combination of cash and the tangency portfolio:

$\mu_p = r_f + S_{\text{tan}}\,\sigma_p.$

This is the capital market line (CML). It is a straight line, and it dominates the risky frontier: at 9.09% volatility, the CML earns about 6.17%, while the 50/50 portfolio earns 6.00% at the same risk. The difference comes from holding the tangency portfolio and levering it slightly with borrowing at the risk-free rate.

The two-fund separation theorem follows: every efficient investor holds the same risky portfolio (the tangency) and only differs in how much cash or leverage to add. In the CAPM, the tangency portfolio is the market portfolio, and the CML becomes the capital market line relating expected return to beta.

With two assets, the tangency weight has a closed form:

$$w_t = \frac{(\mu_A-r_f)\sigma_B^2 - (\mu_B-r_f)\rho\sigma_A\sigma_B} {(\mu_A-r_f)\sigma_B^2 + (\mu_B-r_f)\sigma_A^2 - (\mu_A+\mu_B-2r_f)\rho\sigma_A\sigma_B}.$$

For the example, $w_t = 0.309$. With long-only constraints, the maximum-Sharpe portfolio can be at a corner instead of an interior tangency. With shorting allowed, it can require negative weights. Both are worth checking before treating the tangency as a practical portfolio.

The Playground

The playground below draws the frontier, the minimum-variance portfolio, the maximum-Sharpe portfolio, and the capital market line. The white dot is your current portfolio; the dotted line from cash shows its Sharpe ratio.

A few experiments:

• Drag the weight in A from 0% to 100%. The white dot moves along the curve. The current portfolio turns rose when it is below the frontier.
• Lower the correlation from 0.2 to −0.9. The frontier bends left; the minimum-variance risk falls; the diversification benefit grows.
• Try the stocks & gold preset (low correlation) and the two tech stocks preset (high correlation). The difference in the shape of the frontier is the difference between diversification and decoration.
• Raise the risk-free rate. The CML rotates; the tangency moves; the max-Sharpe portfolio changes. If cash pays more than every risky mix, cash dominates and the CML disappears.
• Watch the diversification readout: it is the gap between the weighted-average volatility and the actual portfolio volatility. At 50/50 in the default example, it is 1.91 percentage points.

The chart uses long-only weights. The tangency formula can produce weights outside $[0,1]$; in that case, the maximum-Sharpe portfolio shown is the best feasible corner, not a true tangency.

Where MPT Breaks

The efficient frontier is a beautiful piece of mathematics, but it is only as good as its inputs.

Expected returns are estimates. They are noisy, unstable, and easy to get wrong. The optimizer is sensitive to them: a small change in $\mu$ can produce a large change in the optimal weights. Michaud called this "error maximization" — the optimizer confidently allocates to whatever input is most overstated.

Correlations are not constants. They tend to rise in crises, exactly when diversification is most valuable. A portfolio that looked diversified in calm markets can behave like one asset in a crash.

Returns are not normal. Mean-variance uses only the first two moments. It ignores fat tails, skewness, and the drawdowns that actually make investors panic.

It is a one-period model. It ignores taxes, transaction costs, rebalancing, liabilities, and the fact that most investors care about the path, not just the endpoint.

Constraints and leverage matter. The tangency may require shorting; long-only portfolios may have a corner solution; borrowing costs more than the risk-free rate; leveraged portfolios can be forced to deleverage at the worst time.

Practical implementations add shrinkage, resampling, Black–Litterman views, factor models, risk budgets, and constraints. These do not refute the frontier; they try to make its inputs less wrong. The frontier is a map of possibilities, not a promise about the future.

MPT, Kelly, and Shannon

These three ideas are often taught separately. They are three layers of the same problem.

• MPT chooses the mix of risky assets. The tangency portfolio is the maximum-Sharpe direction.
• Kelly chooses the scale. The growth-optimal portfolio maximizes expected log wealth. For small returns,

$\log(1 + w^\top R) \approx w^\top \mu - \frac{1}{2}w^\top \Sigma w,$

which is mean-variance optimization with risk aversion 1. With a risk-free asset, the Kelly weights are

$w_{\text{Kelly}} = \Sigma^{-1}(\mu - r_f\mathbf{1}),$

which is a scaled version of the tangency portfolio. The tangency gives the direction; Kelly gives the leverage.

• Shannon's demon is the dynamic version. Rebalancing a volatile portfolio harvests the variance that MPT treats as a static input. The rebalancing premium and the diversification benefit are cousins: both come from imperfect correlation and from the difference between the arithmetic average and the compound path.

In one sentence: MPT says what to own, Kelly says how much, and Shannon explains why rebalancing helps along the way. Kelly sizing is covered in the Kelly criterion article; the dynamic rebalancing story is in the Shannon's demon article.

Conclusion

The efficient frontier is a useful mental model, not a money machine. It says that risk is not a single number to minimize; it is a dimension of choice. It says that correlation is the lever that makes diversification work. And it says that the best portfolio depends on how much risk you are willing to take.

The hard part is not drawing the frontier. It is estimating the inputs well enough that the frontier is not a portrait of your errors. Treat the curve as a map: it shows the terrain, but you still have to choose where to stand.

Further Reading

• Harry Markowitz, "Portfolio Selection" (1952) — the original mean-variance paper
• James Tobin, "Liquidity Preference as Behavior Towards Risk" (1958) — two-fund separation
• William Sharpe, "Capital Asset Prices" (1964) — the CAPM and the capital market line
• Richard Michaud, "The Markowitz Optimization Enigma: Is Optimized Optimal?" (1989) — the estimation-error critique
• Fischer Black & Robert Litterman, "Global Portfolio Optimization" (1992) — combining views with the market
• Richard Grinold & Ronald Kahn, Active Portfolio Management — the practitioner's treatment
• Antti Ilmanen, Expected Returns — a practitioner's survey of risk premia, correlations, and portfolio construction