Why does nobody actually sit on the efficient frontier?
The frontier is a beautiful result and a treacherous tool. The gap between the two is the most professional idea in this course: the curve is drawn with numbers nobody knows.
The three inputs, and their reliability
Drawing a frontier requires three ingredients for the holdings involved — expected returns, volatilities and correlations. All three are estimated from history, and they differ enormously in trustworthiness.
- Volatility is estimated tolerably well from past data. It is persistent: a turbulent asset tends to stay turbulent for a while.
- Correlation is estimated less well, moves around, and — as the next unit shows — shifts exactly when it matters.
- Expected return is the weakest of the three by a wide margin. Distinguishing a 6% long-run expected return from a 9% one with statistical confidence takes decades of data, and the underlying economics can change faster than the data accumulates.
The frontier's position depends most sensitively on the input we know least.
A fourth reliability problem is the price series itself. Listed assets are priced by trades every day. Private equity, property and private credit are priced by appraisal, and an appraisal moves less and later than a market would. The measured volatility of such a series is low, its measured correlation with listed markets is low, and both are properties of the marking, not of the asset. An optimiser fed those inputs treats the unlisted sleeve as the best diversifier available and allocates to it heavily. Before an appraisal-based series enters a frontier, unsmooth it or read its risk from the listed equivalent. A private-equity fund holds companies, so its risk is at least that of small listed companies with leverage.
Error maximisation
Here is the uncomfortable mechanic. An optimiser told to find the best risk-return combination will load up on whichever holdings look best in the input data — and the ones that look best are disproportionately the ones whose expected returns were overestimated. The procedure does not merely tolerate estimation error; it seeks it out and concentrates in it. Practitioners nicknamed naive optimisers "error maximisers," and the label is fair.
The visible symptom: feed in slightly different history and the "optimal" weights swing wildly — 60% in one holding on one run, 5% on the next, from data that barely changed. Output that unstable is telling you about the noise in its inputs, not about the world.
The frontier moves
Even setting estimation aside, the true frontier is not a fixed object. It is drawn from the current joint behaviour of assets, and that behaviour shifts with regimes: the stock-bond relationship of the low-inflation 2010s was not the relationship of 2022. A frontier estimated on one regime describes a world that may no longer exist by the time anyone acts on it.
What professionals actually keep
They keep the logic and distrust the coordinates. The durable lessons survive the estimation problem entirely:
- Dominated combinations exist, and holding one is a genuine inefficiency.
- Lower correlations expand the opportunity set — that is structural, not estimated.
- Concentration in a single view is fragile whether or not an optimiser says so.
What does not survive is the illusion of a precise optimal point. The honest posture is the one the Foundations courses trained: use the model to organise thinking, never to manufacture certainty. A frontier is a hypothesis about the future, drawn with a ruler made of the past.
In the data
Nobody publishes a covariance matrix or a frontier as market data. What is public is the price history they are estimated from, here the two funds behind the classic stock-bond frontier:
Every estimate starts with two choices made on lines like these: which years, and daily or monthly bars. The same two funds over the same dates give a different covariance from daily bars than from monthly ones, and a different one again from the next five years. The optimiser's output carries no record of which it was fed.
Try it now
Estimate one pair's co-movement three times from the two charts above. First over 2013–2017, then over 2018–2022 — same pair, different five years. Then switch both charts to Monthly and do one of those windows again. Three estimates, one pair of instruments, and an optimiser's output carries no record of which of the three it was fed.
Take a holding you follow and try to state its expected annual return over the next decade to within one percentage point. Notice how uncomfortable the exercise is — that discomfort is the frontier's weakest input, and it is the one the optimiser is most sensitive to.
Write the caveat in your own words: the frontier's shape is a durable insight; its exact coordinates are an estimate.