Variance is the wrong definition of risk

In Part 1 we covered the classical mean-risk family, all built on variance as the measure of risk. But variance has a strange property: it penalizes upside deviation exactly as much as downside. A portfolio that occasionally surprises to the upside is treated as just as "risky" as one that occasionally collapses. No real investor feels that way. What keeps people up at night is not volatility — it is loss, and especially deep, sustained loss.

This installment covers the families built for that intuition: methods that optimize against tail losses and drawdowns, methods that allocate by risk contribution rather than return, and methods that use the correlation structure of the market itself. As in Part 1, every output here is reviewed research — a candidate for portfolio, tax, and compliance review, not an order.

Tail risk: optimizing for the bad days

The first step beyond variance is to focus on the tail of the distribution — the rare, severe outcomes:

• Minimum CVaR — minimize the average loss in the worst slice of outcomes (say, the worst 5%). Where variance asks "how much does this bounce around?", CVaR asks "when it goes wrong, how wrong?"
• Minimum EVaR — an entropic value-at-risk measure that penalizes severe tail outcomes more conservatively.
• Worst-Case Realization (Minimax) — optimize against the single worst modeled outcome. Deliberately pessimistic, and scenario-dependent.
• Distributionally Robust CVaR — extend CVaR to acknowledge you are not even sure the return distribution itself is right.

The review focus across these is the same: how deep is the tail you are modeling, how much expected return are you sacrificing for that protection, and is the chosen measure something you can actually explain to the person whose money it is? A tail measure nobody understands is a liability dressed as prudence.

Drawdown: the risk you actually live through

Tail risk looks at point-in-time losses. But investors experience risk as a path — the agony of watching a portfolio sit underwater for months. Drawdown methods optimize that lived experience directly:

• Minimum CDaR / EDaR — minimize conditional (or entropic) drawdown at risk rather than volatility.
• Minimum Maximum Drawdown — attack the single largest peak-to-trough fall.
• Minimum Ulcer Index — penalize both the depth and the persistence of drawdowns, which captures how it actually feels to stay underwater.

These shine for drawdown-sensitive mandates — anyone drawing income, anyone near retirement, anyone who might capitulate at the bottom. The key review question is whether the historical window used is representative; a drawdown method tuned to a calm decade will be unprepared for a turbulent one.

Risk parity: allocate risk, not dollars

A different idea entirely: stop allocating by expected return — which we know is hard to forecast — and allocate by risk contribution instead.

• Risk Parity (Equal Risk Contribution) — size positions so each contributes the same share of total portfolio risk. A small bond sleeve and a volatile equity sleeve get balanced by their risk, not their dollar weight.
• Risk Budgeting on CVaR / EVaR / CDaR — distribute tail or drawdown risk budgets rather than variance contributions.
• Relaxed Risk Parity — loosen strict parity when the rigid version is too constraining.

The appeal is that risk contribution is more stable and estimable than expected return. The watch-outs are leverage (true risk parity often assumes you can lever the low-risk sleeve) and concentration (low-volatility assets can quietly dominate the allocation).

Hierarchical methods: respect the market's structure

Classical optimizers treat the covariance matrix as one big block and often produce unstable results because that matrix is noisy and hard to invert. Hierarchical methods take a smarter route — they first discover the structure of the market:

• Hierarchical Risk Parity (HRP) — cluster correlated assets into a tree, then allocate down through the tree rather than inverting one giant matrix.
• Hierarchical Equal Risk Contribution (HERC) — combine that clustering with equal-risk-contribution allocation.
• Nested Cluster Optimization — optimize within clusters, then across them.
• Schur Complementary Allocation — decompose covariance to separate common from residual risk.

By respecting how assets actually group together, these methods tend to produce more stable, intuitive allocations that survive better out of sample. The review focus is the cluster structure itself — the dendrogram, the linkage method, and whether any single cluster quietly concentrates the portfolio.

Choosing among them

These families are not competitors so much as different lenses. A drawdown- sensitive retiree, a tail-averse institution, and a manager who distrusts return forecasts will each gravitate to a different one — and the right way to choose is to run two or three against the same portfolio and compare them on the metrics that matter: tail loss, maximum drawdown, risk contribution, turnover, and what each gives up in expected return. The comparison is the decision aid; the suitability and execution review still come after.

The takeaway

Once you accept that variance is not the same as risk, a whole landscape of optimizers opens up: tail-focused methods for severe losses, drawdown methods for the path you actually live through, risk parity for allocating risk instead of guessing returns, and hierarchical methods that respect the market's real structure. Each is built for a specific fear. In Part 3 we close the series with the methods that incorporate your views and your constraints — Black-Litterman, factor models, turnover and tax budgets, and the robust optimizers built for a world where your estimates are wrong.