The Efficient Frontier
The set of portfolios offering the highest expected return for each level of risk. A foundational idea, and one whose inputs are estimates with enough error to move the answer entirely.
MadStockAlerts Research · Updated September 23, 2026
What to Take Away
- The frontier is derived from expected returns, volatilities and correlations.
- All three are estimates, and expected returns are the least reliable of them.
- Small changes in inputs produce large changes in the optimal weights.
- Optimizers concentrate into whatever asset has the most favorable estimate.
- The framework's durable contribution is the idea that risk is a portfolio property.
MAD Academy Training Video · 0:45
An Idea Worth Knowing, and Its Limits
The frontier shows the best return available at each level of risk, computed from inputs nobody can actually know in advance.
This lesson is part of a MAD Academy plan.
The construction
Given a set of assets, each with an expected return and a volatility, and a matrix of correlations between them, it is possible to compute the combination that produces the highest expected return for any given level of volatility. The set of those combinations is the efficient frontier.
The insight that made this influential is that a portfolio's risk is not the average of its components' risks. Combining imperfectly correlated assets produces a portfolio whose volatility is lower than the weighted average of theirs, and that gap is the mathematical content of diversification.
That contribution is durable and is why the framework is taught. An asset is risky in the context of what it is held alongside, not on its own, and this is where that idea was formalised.
The estimation problem
| Input | How it is estimated | Reliability |
|---|---|---|
| Volatility | From historical returns | Reasonably stable, and it clusters |
| Correlation | From historical returns | Unstable, and it changes in stress |
| Expected return | From history, or a model, or judgment | Very poor. This is the binding problem |
The third row is where the framework meets its limit. Expected returns cannot be estimated from history with useful precision: the standard error on a mean return estimated from decades of data remains large enough to encompass most plausible values.
The output is highly sensitive to exactly that input. A change in an expected return assumption of half a percentage point can shift an optimized allocation by tens of percentage points, which means the answer is largely a restatement of the assumption.
Scroll the chart sideways to see all of it.
- Expected return for A: 7.0%
- Expected return for A: 7.5%
Error maximization
An optimizer allocates toward whatever combination of inputs looks best. Where an input is overestimated, the optimizer responds by allocating more to it, which means estimation errors are amplified rather than averaged out. The behavior is sometimes described as error maximization.
- Unconstrained optimization typically produces extreme, concentrated weights.
- Adding constraints produces more reasonable portfolios and does so by overriding the optimization.
- Techniques exist to address this, including shrinkage of the estimates and resampling, and they reduce rather than remove the problem.
- The result is that most practical allocation work uses the framework as a way of thinking rather than as a calculator.
What survives
Setting aside the optimizer, several conclusions from the framework have held up well and are worth separating from the machinery.
- Risk should be assessed at the portfolio level rather than position by position.
- Combining imperfectly correlated assets improves the return available per unit of risk.
- Adding a volatile asset can reduce total portfolio risk if its correlation is low enough.
- Concentration has a cost that is measurable, even where the concentrated asset is a good one.
The third item is the counterintuitive one and it is genuinely true. It is also the point at which the framework's dependence on a correlation estimate becomes most consequential.
The practical alternatives
Because optimization amplifies estimation error, several approaches deliberately use less information, and they have generally performed at least as well out of sample.
| Approach | What it uses | Why it works |
|---|---|---|
| Equal weight | Nothing but the asset list | No estimates to be wrong about |
| Risk parity | Volatilities and correlations, not expected returns | Drops the least reliable input entirely |
| Minimum variance | The covariance matrix only | Same reasoning, optimizing a different objective |
| Constrained optimization | All three, with limits on weights | The constraints prevent the extreme allocations |
The first row is not a joke. Studies comparing equal weighting against optimized portfolios out of sample have repeatedly found the simple approach competitive, which is a statement about how poor the expected-return estimates are rather than about the mathematics.
The general lesson generalizes past allocation: where an input cannot be estimated reliably, a method that does not require it is frequently better than a better method that does.
Finished this lesson?
Progress tracking, subject trophies and the training videos are part of every paid plan.