The efficient frontier explained: why correlation matters more than counting investments

Efficiency means a comparison within a feasible set

A mean–variance efficient portfolio offers the highest expected return available for a given variance, or the lowest variance for its expected return, within the choices allowed by the model. A dominated portfolio has an available alternative with a more favourable combination. Neither term means that the investment is safe, suitable or certain to earn its expected return.

Markowitz’s portfolio-selection paper provides the foundational framework. In the efficient frontier simulator, two hypothetical assets have expected returns of 5% and 9% and standard deviations of 10% and 20%. Weights are nonnegative and sum to one. These inputs are supplied teaching assumptions, not estimates of named investments. The displayed frontier is therefore conditional on a very small feasible set and a particular definition of risk.

The covariance term changes the shape of the opportunity

Expected portfolio return is a weighted average of the two supplied means. Variance also includes covariance: the way the two returns move together. With imperfect correlation, some fluctuations offset, so a mix can have less volatility than either asset alone. This benefit depends on the weights and relationship between returns rather than on a generic label such as diversified.

The board lets you change correlation while leaving individual means and standard deviations fixed. Watch the curve move and reveal the minimum-variance point. On one side of that point, a portfolio can have lower expected return and higher risk than another available mix. On the efficient branch, moving toward higher expected return requires accepting more modelled volatility. The calculation describes that tradeoff without deciding how a particular reader should value it.

Worked example: inspect a mix instead of assuming a benefit

Set correlation to zero and allocate thirty percent to the second asset. The following arithmetic comes entirely from the board’s invented moments. It illustrates covariance accounting rather than a forecast of realised investment performance.

1. Compute expected return as 0.7 times five plus 0.3 times nine, giving 6.2 percent. The weighted mean is straightforward, but it does not determine the portfolio’s standard deviation.
2. Calculate variance in squared-percentage units as 100 times 0.7 squared plus 400 times 0.3 squared. With zero correlation the cross term vanishes, giving eighty-five and a standard deviation of about 9.22 percent.
3. Change correlation to one while keeping the same weights. Standard deviation becomes thirteen percent. The same two assets and allocation now supply less diversification because the assumed return movements no longer offset.

A clean curve can conceal uncertain inputs

Real expected returns and covariances are not printed on assets. They must be estimated or assumed, and the answer can change when the period, data quality or market conditions change. A mathematically efficient point under one set of inputs can be inefficient under another. The precision of the plotted curve should not be confused with confidence in those inputs.

The diversification experiment is useful for exploring co-movement separately from a recommendation. Ask what happens if correlations rise during a period when protection is most valuable. Also ask what standard deviation misses: losses can be asymmetric, trading can become costly and a commitment may be hard to unwind. Adding these dimensions changes the decision problem rather than merely adjusting a label on the same chart.

Five questions to attach to a frontier chart

A useful frontier graphic should be accompanied by enough information to reconstruct its meaning. Without that information, readers can easily compare points generated from different assets, time horizons or constraints as though they occupied one shared opportunity set. Use the following questions to turn the picture into an inspectable model rather than an allocation instruction.

• Which assets, weights and restrictions define the feasible portfolios being compared?
• Are expected returns assumptions, historical averages or estimates from a stated forecasting model?
• What data and time horizon support the covariance and correlation inputs?
• Which risks, fees, taxes and liquidity constraints are omitted from the variance measure?
• How does the apparently preferred point change when important assumptions move together?

One risk number need not capture a nonlinear exposure

A portfolio can also contain exposures whose response to market changes is curved. For a simple option-free bond, price is a nonlinear function of yield. Duration approximates the local slope; convexity adds a second-order correction. A single sensitivity evaluated at one point is not an exact description of every possible rate move.

Explore this distinction in the bond convexity board, which compares exact discounting with linear and quadratic estimates for fixed cash flows. University of Virginia’s duration and convexity note provide university context on interest-rate sensitivity. This extension does not invalidate mean–variance analysis; it identifies another modelling layer. First define the exposure and scenario, then decide which approximation is adequate. A useful model is one whose limits remain visible when the decision becomes more demanding.