Guide · Part II — The mental tools

Markowitz and broken correlations

Diversification is the only widely available tool that reduces risk without reducing expected return — and it switches off at the exact moment it is needed most. Both halves of that sentence are true, both are measurable, and holding them together is what separates using Markowitz from being used by it.

The problem

An allocator holding one asset carries all of that asset’s risk. Common sense says spreading across several assets helps, but before 1952 nobody had said precisely how much, or shown that the mix could be engineered. Harry Markowitz’s insight was that a portfolio’s risk depends not just on each asset’s volatility but on how the assets move together — the correlations — and that mixing imperfectly correlated assets produces a whole calmer than its parts.

Given a menu of assets, there is a mix that delivers any target return at the lowest possible volatility. The line of those best mixes is the efficient frontier; any portfolio below it takes needless risk for its return. This is the chassis most institutional portfolios are still bolted onto, seventy years on.

The insight

The frontier is only as good as its two inputs, and both are guesses. Expected returns are backward-looking or modelled, always uncertain — and a 1% error in one asset’s estimate can swing its “optimal” weight by 20 to 30 percentage points. Optimisers are, in the trade phrase, error maximisers: they pour weight into exactly the assets whose rosy estimates are most likely wrong.

The second input is the correlation matrix, and it is less stable than it looks. Twenty assets means 190 pairwise correlations to estimate, and all of them shift under stress. Correlations are measured mostly in calm markets, because most history is calm — and in a panic, frightened holders sell everything at once to raise cash, so assets that spent a decade moving independently fall together.

The umbrella folds when it starts to rain.

2022 is the textbook case. The stock-bond correlation, negative for two decades, flipped positive in an inflation shock; “balanced” 60/40 portfolios — sold for a generation as automatically safe — behaved in realised risk almost like all-equity books. Illustratively, equities down around 18% and “safe” bonds down around 13% blend to roughly minus 16%, with the protective half doing almost none of its job. The model assumed the diversifier would diversify; the regime decided otherwise.

Two further failure modes complete the set. Markowitz assumes returns follow the bell curve, and real returns have fat tails — extremes far more frequent than the Gaussian allows, so the frontier under-sizes disaster for any asset whose defining feature is violent moves. And the model is regime-blind: a frontier optimised on risk-on data is dangerous in risk-off weather, which is why serious practitioners recompute per regime rather than once.

In plain English

Markowitz draws a flawless map of a calm harbour. In a storm, all the boats start moving together, and one or two of the boats turn out to be fireworks. The map remains genuinely useful — for calm days, with boats whose behaviour is understood — but it must never be mistaken for a promise about storms.

Three working rules fall out.

• Seatbelt, not airbag. Treat diversification as the everyday seatbelt: it works in normal bumps and disengages in the crash, so no position should be sized as if its diversification will hold in a panic.
• Start at minimum variance. Prefer the minimum-variance end of the frontier as a starting point, because it leans least on the fragile return estimates.
• Cap every position. Cap any single position — a common institutional rule is 20% — as a hard, unconditional limit that blunts the error-maximiser pathology without requiring the inputs to be right.

There is also a quieter test hiding here: whether “different” holdings are actually different. A hypothetical Singapore portfolio holding local REITs and local blue-chip banks is not two bets — both ride the same domestic rate curve and the same handful of lenders. Diversification across names is not diversification across risks; the correlation question has to be asked at the level of what actually drives each holding.

Where this breaks

Beyond the crisis-correlation failure already described, mean-variance has a structural blind spot: it optimises inside a measuring unit it never questions. The mathematics is anchored to a “risk-free” rate on government debt, and in periods of sustained currency debasement that anchor carries a guaranteed real loss. There is no slot in the model for “the ruler is shrinking.” Nor is there a slot for binary events — default, reclassification, custody failure — that arrive as a jump rather than a drift.

Used honestly, Markowitz is a discipline for the middle of the distribution, paired deliberately with tools built for the tails. The next two chapters supply those tools: Kelly for how much to bet, and CVaR for how bad the bad case really is. Before moving on, one exercise: for any two holdings believed to be diversifying each other, name the single event that would make them fall together. If the event is easy to name, the diversification is conditional — size accordingly.