Portfolio risk: what the weights conceal

A portfolio is not a collection of isolated bets

Two positions with identical standalone volatility can contribute very different amounts of risk to an existing portfolio. The difference comes from how their returns interact with other holdings. A position that usually moves with the largest exposure can amplify its fluctuations; another may offset some of them. Counting symbols does not measure this relationship.

The classical mean-variance framework describes portfolio variance using weights, individual variances and pairwise covariances. This is a useful organizing model, not a complete definition of danger. It does not directly capture default, market closure, withdrawal restrictions or permanent loss of purchasing power. Begin by deciding which risks the calculation can illuminate and which require separate analysis.

Make the cross-term visible

For two assets with weights w1 and w2, volatilities s1 and s2, and correlation r, variance is w1²s1² + w2²s2² + 2w1w2rs1s2. The final term is why simply averaging volatilities usually fails. All inputs must use consistent time horizons, return conventions and currency treatment.

An equally weighted pair of assets, each with 20% annualized volatility, has approximately 14.14% modeled volatility when correlation is zero. With correlation one, the result is 20%. With correlation minus one and identical magnitudes, the idealized variance is zero. That last calculation demonstrates the model's arithmetic, not the availability of a real-world riskless portfolio.

Correlation is an estimate, not a contract

A sample correlation changes with the observation window and market environment. Quiet periods can conceal common vulnerability to inflation, funding shortages or forced selling. Assets that look unrelated on ordinary days may react together when investors need cash. Estimation uncertainty should therefore be visible rather than buried inside a single matrix.

Compare several plausible correlations and inspect specific stress episodes. Be careful with stale prices: an infrequently valued asset can appear artificially stable and weakly correlated simply because its recorded value adjusts late. Align observation times and understand valuation methods before interpreting a low number as diversification. Different trading hours create a similar problem even when both instruments are actively traded.

Covariance inputs must describe a mathematically compatible set of relationships. Pairwise correlations estimated from different date subsets can produce a matrix that is not positive semidefinite, meaning some weighted combination would have a negative calculated variance. That is a data or modeling problem, not an unusually powerful hedge. Using aligned observations, checking the matrix’s numerical properties and documenting any regularization are part of the calculation. With many assets and little history, a technically valid sample matrix can still be very unstable.

Translate names into economic exposures

A technology stock fund, an individual software company and a growth-sensitive currency position can share sensitivity to risk appetite even though they have different labels. A domestic account holding foreign assets also carries currency exposure unless it is actually hedged. Decompose positions into drivers such as equity beta, interest rates, currencies, credit and commodity prices.

Derivative notional and cash allocation need special care. A small margin deposit can support a large economic exposure, while an option's sensitivity changes with price and time. Portfolio weights based only on cash paid can seriously misrepresent these instruments. A first-order factor view is useful, but nonlinear exposures still need revaluation under scenarios rather than a fixed sensitivity forever.

Use marginal changes to inform analysis

Ask what happens to total modeled risk if a small amount of a position is added or removed. This marginal perspective can reveal that the largest holding is not necessarily the largest incremental risk. It also prevents a standalone volatility ranking from being mistaken for a portfolio decision.

Keep the covariance calculation alongside concentration limits, liquidity measures and loss scenarios. A mathematically low-volatility mix may contain a single shared custodian or depend on leverage that cannot survive stress. The practical output should be a map of vulnerabilities and uncertainties, not one decimal place of apparently definitive risk. Recalculate after meaningful exposure changes and investigate why the result moves.

Make exposure and funding visible in the weights

Weights need a denominator. A $60,000 long position and a $40,000 short position in an account with $100,000 equity have equity-relative weights of +0.60 and −0.40. Their net exposure is 20% of equity, but their gross exposure is 100%. Renormalizing the two risky positions to sum to one would silently change their economic scale. Cash, short-sale proceeds and financing must be accounted for separately under the actual arrangement, rather than inferred from a tidy list of weights.

For those two positions, assume each asset has 20% annualized volatility and their return correlation is 0.8. Using the signed weights gives variance of 0.60² × 0.20² + (−0.40)² × 0.20² + 2 × 0.60 × (−0.40) × 0.8 × 0.20 × 0.20 = 0.00544. Modeled volatility is about 7.38% of account equity, ignoring variable financing costs. If correlation falls to zero, modeled volatility rises to about 14.42%. Positive correlation helps this long-short hedge; it hurts an otherwise comparable pair of long positions.

Currency conversion creates another interaction. If a foreign asset gains 5% in local currency while that currency falls 10% against the account currency, the account-currency return is 1.05 × 0.90 − 1 = −5.5%, before costs. Adding 5% and −10% gives only an approximation because it omits the cross-product. Calculate the return series in the currency relevant to the account before estimating portfolio covariance, unless a separately modeled hedge changes the exposure.

Finally, distinguish a fixed-weight model from an actual rebalancing policy. Price moves change weights unless trades restore them. A covariance estimate made at today’s weights describes local exposure; it does not automatically describe next month’s drifting portfolio or the costs of daily rebalancing. Compare the intended policy with feasible execution, and include the funding and liquidity consequences that remain invisible in the variance calculation.