Correlation and Risk Concentration: Count Exposures, Not Labels

Distinguish association from independence

NIST defines the correlation coefficient as a measure of the linear relationship between two variables. Its range runs from minus one to plus one when the required variances are nonzero. A value near zero says little about some nonlinear relationships. It also says nothing by itself about the size of the underlying changes. Low correlation between two very volatile exposures can still produce substantial portfolio variability.

The measured variables matter. Correlation between raw reference-price changes is not necessarily correlation between a strategy's signed, resized exposures. A rule can change direction, move to neutral or adjust size. Before interpreting a matrix, identify whether it describes prices, comparable returns, hypothetical dollar changes or strategy outcomes. A matrix with unclear inputs may look sophisticated while answering the wrong concentration question.

Calculate a two-exposure example

Imagine two hypothetical sleeves, each with annual volatility of 20%, assigned equal 50% capital weights. Portfolio variance is the sum of the two weighted variance terms plus twice the weighted covariance. With zero correlation, the calculation is 0.5 squared times 0.2 squared, repeated twice: variance equals 0.02, and portfolio volatility is approximately 14.1421%. This is conditional arithmetic, not a forecast for two real funds.

If correlation is 0.8, the covariance term adds 0.016, giving variance of 0.036 and volatility of 18.9737%. At correlation one, volatility is 20%. MIT's portfolio-theory materials explain this covariance-based framework. The example shows that the apparent diversification benefit can shrink without either sleeve changing its own estimated volatility or capital weight. The relationship between them is itself a major input.

Position signs can reverse the interpretation

Suppose two underlying return series have a positive correlation of 0.8. A fixed long position in the first and an otherwise comparable fixed short position in the second reverse the sign of their covariance contribution. For this simplified pair, multiplying one series by minus one changes their correlation to minus 0.8. Reading the unsigned underlying matrix alone would therefore misdescribe the relationship between these signed exposures.

That observation does not make every long-short pair safe. Different volatility scales, imperfect matching, currency effects and changing positions can weaken the offset. If the relationship shifts, the hedge can behave differently from the estimate. Examine the actual exposure vector used with the covariance matrix, and avoid summarising a changing strategy with a correlation measured from unrelated fixed-direction price series.

Capital weights do not equal risk weights

Consider another fictional pair with equal capital weights, zero correlation and individual annual volatilities of 10% and 30%. Their weighted variance terms are 0.0025 and 0.0225. The second contributes 90% of total variance under these particular assumptions, despite receiving only half the capital. Total variance is 0.025, giving portfolio volatility of 15.8114%. Equal money allocations have not created equal risk contributions.

When correlation is nonzero, allocating covariance terms introduces further choices about attribution. Be clear whether a report shows standalone risk, marginal contribution or a scenario loss. Those labels are not synonyms. A practical concentration review begins by asking which exposure dominates the chosen measure, rather than assuming the largest capital weight or greatest contract count must be the largest source of risk.

Audit the sample behind the matrix

Align observations on a meaningful common calendar. A close recorded in one time zone may represent a different information set from another market's close on the same printed date. Missing observations and stale prices can make co-movement appear weaker than it is. State the alignment convention, return horizon and sample endpoints before drawing conclusions from a low historical coefficient.

Using different available date pairs for every correlation can also produce a matrix that is not internally coherent for portfolio calculations. A common valid sample is easier to explain, although it can discard information. Whatever method is chosen, disclose it and check the matrix's mathematical validity. A negative calculated portfolio variance is an implementation or input-consistency problem, not evidence of exceptionally effective diversification.

Add driver and stress views

A statistical matrix should be accompanied by a plain-language map of possible shared drivers. Several positions might respond similarly to financing conditions or a common currency even when their product labels differ. These are hypotheses to examine, not permanent classifications. The objective is to identify plausible ways positions could move together that a quiet or short sample may not have represented.

Recalculate the fictional equal-volatility pair under the alternative correlations in the table without attaching invented probabilities to them. Separately consider simultaneous adverse moves and reduced liquidity. Correlation stress is only one dimension: increasing both standalone volatility and co-movement can produce a larger change than either alone. A compact scenario grid makes those dependencies visible without pretending to know the next market regime.

Make concentration limits interpretable

Limits can be expressed by gross exposure, estimated risk contribution, common-driver group or a specified scenario loss. Each needs a clear denominator and a policy for conflicting signals. A portfolio can pass a position-count limit yet fail a common-driver stress. Conversely, a net exposure near zero can coexist with large gross positions that demand collateral and generate turnover during rebalancing.

Keep estimates and decisions separate in the record. Show the observation window, signed weights, individual scales and scenario assumptions that led to the concentration assessment. The useful conclusion is not that a particular number of holdings guarantees diversification. It is that the portfolio has been examined as a combination of exposures, including the possibility that relationships change when the hoped-for offsets matter most.

Source record

1. NIST Dataplot: Correlation
2. MIT OpenCourseWare: Portfolio Theory