Portfolio Risk and the Incentive to Diversify
A risk measure does more than assign a number to a loss: when applied consistently, it also affects how a firm evaluates combining exposures. The key comparison is between the risk measure of the portfolio as a whole and the sum of the measures assigned to its components. If the portfolio measure is lower, the measure records a diversification benefit. If it is higher, combining the exposures attracts a measured penalty.
This comparison is about the incentives created by the risk measure, not a claim that every real-world merger of exposures improves outcomes. The losses being compared must use the same time horizon, valuation basis, and loss convention. Their dependence also matters: the combined loss is the random variable \(X+Y\), whose distribution depends on the joint behavior of \(X\) and \(Y\).
A coherent risk measure is monotone, translation equivariant for losses, positively homogeneous, and subadditive. In particular, subadditivity says that \(\rho(X+Y)\leq\rho(X)+\rho(Y)\). That inequality rules out a measured diversification penalty. It does not promise a strictly positive benefit: equality is possible.
What the Axioms Guarantee—and What They Do Not
The incentive implication follows directly from the definition of \(D_{\rho}\). Under subadditivity, the combined portfolio cannot receive a higher risk assessment than the sum of the stand-alone assessments. Thus a decision-maker using a subadditive measure does not face an axiomatic penalty merely for pooling two losses.
This is a bound, not a prediction of how large the benefit will be. If losses are perfectly aligned, or if the chosen measure responds only to their means, the risk measure may give equality. If dependence is favorable, or if a measure responds to tail behavior that is reduced by pooling, the benefit may be positive. Those outcomes depend on the losses and the particular measure; subadditivity alone establishes only the nonnegative-benefit bound.
A measure that is not subadditive has no such general guarantee. It can still give a lower risk for a particular combined portfolio, but there may be portfolios for which it reports more risk after combination. That distinction is important: failure of subadditivity permits a penalty; it does not mean every portfolio is penalized.
Worked Example: Discrete VaR Can Penalize Pooling
Worked Example: Discrete VaR Can Penalize Pooling
A hypothetical insurer considers two business lines over the same one-year period. Loss \(X\) is \$100,000 if event A occurs and zero otherwise. Loss \(Y\) is \$100,000 if event B occurs and zero otherwise. Events A and B are mutually exclusive, each with probability 0.04; neither event occurs with probability 0.92. Calculate the 95% VaR of each loss and of the combined loss, then assess the diversification incentive. Use \(\operatorname{VaR}_{\alpha}(X)=\inf\{x:F_X(x)\geq\alpha\}\).
Method. Find each loss distribution’s cumulative probability at zero, then apply the stated quantile definition at \(\alpha=0.95\).
For \(X\), the probability of zero loss is \(1-0.04=0.96\). Therefore \(F_X(0)=0.96\geq0.95\), so \(\operatorname{VaR}_{0.95}(X)=\$0\). The same calculation gives \(\operatorname{VaR}_{0.95}(Y)=\$0\).
For \(X+Y\), mutual exclusivity means the total is \$100,000 when either event occurs, with probability \(0.04+0.04=0.08\), and zero otherwise. Thus \(F_{X+Y}(0)=0.92<0.95\), while \(F_{X+Y}(\$100{,}000)=1\geq0.95\). Hence \(\operatorname{VaR}_{0.95}(X+Y)=\$100{,}000\).
The measured diversification benefit is \(D_{\operatorname{VaR}}=\$0+\$0-\$100{,}000=-\$100{,}000\). This VaR calculation assigns a larger risk to the pooled losses than to the sum of the stand-alone VaRs. Common mistake: Adding the two stand-alone VaRs and calling the result the portfolio VaR. That sum is \$0 here, but the portfolio quantile must be calculated from the combined distribution; it is \$100,000.
Check: The 95th percentile of a loss that is zero with probability 0.92 and \$100,000 with probability 0.08 is \$100,000 under the stated infimum definition. The result is consistent with the cumulative distribution crossing 0.95 at \$100,000.
The example shows why a candidate must check the distribution of the aggregate rather than assume that stand-alone tail quantiles add. Discrete probability at a quantile boundary can matter: here, each individual loss exceeds zero only 4% of the time, but the combined loss is positive 8% of the time.
Why Normal TVaR Gives a Different Portfolio Comparison
For a continuous normal loss, TVaR has a convenient form. Let \(X\sim N(\mu,\sigma^2)\), with \(\sigma>0\), and let \(z_{\alpha}\) satisfy \(\Phi(z_{\alpha})=\alpha\), where \(\Phi\) and \(\phi\) are the standard normal distribution and density functions. For this continuous model, TVaR at level \(\alpha\) is the mean loss in the upper \(1-\alpha\) tail.
To see why, standardize the loss as \(X=\mu+\sigma Z\), where \(Z\sim N(0,1)\). The tail begins at \(Z=z_{\alpha}\). The expected standardized loss conditional on exceeding that threshold is \(\frac{1}{1-\alpha}\int_{z_{\alpha}}^{\infty}z\phi(z)\,dz\). Since \(\phi'(z)=-z\phi(z)\), integrating gives \(\int_{z_{\alpha}}^{\infty}z\phi(z)\,dz=\phi(z_{\alpha})\). Rescaling by \(\sigma\) and adding \(\mu\) gives the displayed formula. This form assumes a continuous normal loss distribution; it is not a formula for arbitrary discrete losses.
Worked Example: Independent Normal Losses Under TVaR
A hypothetical portfolio has independent annual losses \(X\sim N(10,8^2)\) and \(Y\sim N(20,6^2)\), with all amounts in thousands of dollars. Calculate the 95% TVaR of each loss and of the combined loss, then find the measured diversification benefit. Use \(z_{0.95}=1.6449\) and \(\phi(z_{0.95})/(1-0.95)=2.0627\).
Method. Apply the normal TVaR formula to each loss. For the aggregate, independence gives \(\operatorname{Var}(X+Y)=\operatorname{Var}(X)+\operatorname{Var}(Y)\), so its mean is \(10+20=30\) and its standard deviation is \(\sqrt{8^2+6^2}=\sqrt{100}=10\).
The stand-alone calculations are:
Their sum is \(26.5016+32.3762=58.8778\) thousand dollars. For the aggregate,
Thus \(D_{\operatorname{TVaR}}=58.8778-50.6270=8.2508\) thousand dollars, or \$8,250.80. The measure records a positive benefit. Common mistake: Adding the standard deviations to get \(8+6=14\). Under independence, the aggregate standard deviation is \(\sqrt{8^2+6^2}=10\), not 14.
Check: The aggregate mean is 30, and its tail uplift is 20.6270; together they give 50.6270. This is below the stand-alone sum of 58.8778, as required by subadditivity.
Subadditivity Does Not Require a Strict Benefit
The comparison also needs a reverse caution: a coherent measure need not reward every combination with a strictly lower risk number. For example, the expected-loss measure \(\rho(X)=E[X]\), for losses with finite expectations, is additive: \(E[X+Y]=E[X]+E[Y]\). It is coherent, but its measured diversification benefit is always zero. Its value does not vary with dependence when the marginal expected losses are fixed.
The following example shows the same equality under TVaR when two normal losses share exactly the same underlying risk factor. The calculation is a case of no measured benefit, not a violation of subadditivity.
Worked Example: A Coherent Measure with No Strict Benefit
Suppose two hypothetical annual losses, in thousands of dollars, are \(X=10+4Z\) and \(Y=20+3Z\), where \(Z\sim N(0,1)\). Calculate their individual and aggregate 95% TVaRs using the normal formula and the tail coefficient 2.0627.
Method. Each loss is normal because it is an affine function of a standard normal variable. Since both coefficients of \(Z\) are positive, the aggregate is \(X+Y=30+7Z\).
The individual means and standard deviations are \((10,4)\) and \((20,3)\). Therefore \(\operatorname{TVaR}_{0.95}(X)=10+4(2.0627)=18.2508\), and \(\operatorname{TVaR}_{0.95}(Y)=20+3(2.0627)=26.1881\). Their sum is \(44.4389\) thousand dollars.
The aggregate has mean 30 and standard deviation 7, so \(\operatorname{TVaR}_{0.95}(X+Y)=30+7(2.0627)=44.4389\) thousand dollars. The benefit is \(44.4389-44.4389=0\). Common mistake: Concluding that a coherent measure must show a strict reduction. Subadditivity allows equality, as this shared-factor portfolio demonstrates.
Check: Because the losses move together in the same direction, the aggregate standard deviation is \(4+3=7\); the aggregate tail uplift \(7(2.0627)=14.4389\) equals the sum of the two individual tail uplifts.
When a Risk Measure Penalizes Combining Exposures
Variance illustrates a different issue. Although useful for describing dispersion, variance is not a coherent monetary risk measure and can be non-subadditive. For positively dependent losses, the covariance term can make aggregate variance exceed the sum of individual variances:
If \(\operatorname{Cov}(X,Y)>0\), the aggregate variance is larger than the sum of the stand-alone variances. There is also a units issue: if losses are measured in dollars, variance is measured in dollars squared. It should not be interpreted directly as a dollar capital requirement.
Worked Example: A Variance-Based Aggregation Penalty
Let \(Z\) be a Bernoulli random variable with \(P(Z=1)=P(Z=0)=0.5\). Two hypothetical loss exposures, in dollars, are \(X=10Z\) and \(Y=10Z\). Calculate each variance and the variance of the combined loss.
Method. For a Bernoulli variable with success probability 0.5, \(E[Z]=0.5\), \(E[Z^2]=0.5\), and \(\operatorname{Var}(Z)=E[Z^2]-E[Z]^2=0.5-0.25=0.25\).
Since \(X=10Z\), \(E[X]=10(0.5)=5\) and \(E[X^2]=100(0.5)=50\). Hence \(\operatorname{Var}(X)=50-5^2=25\) dollars squared. The same calculation gives \(\operatorname{Var}(Y)=25\) dollars squared. The stand-alone variance sum is \(25+25=50\).
The combined loss is \(X+Y=20Z\). Its mean is 10, its second moment is \(400(0.5)=200\), and its variance is \(200-10^2=100\) dollars squared. Thus the aggregate variance exceeds the stand-alone sum by \(100-50=50\) dollars squared. Common mistake: Treating this difference as a \$50 loss or capital amount. It is 50 dollars squared, so it cannot be read as a dollar risk charge.
Check: The two losses are perfectly positively dependent, and \(\operatorname{Cov}(X,Y)=25\). The variance identity gives \(25+25+2(25)=100\), matching the direct calculation.
Interpreting the Incentive Carefully
A coherent measure’s subadditivity supplies a useful discipline for portfolio aggregation: the measure will not report a higher combined risk than the sum of the component risks. A non-subadditive measure lacks this guarantee, so its aggregation result should be examined for potential penalties. Neither statement, by itself, establishes the best measure for a particular decision.
- Compare like with like. Use a consistent loss definition and horizon for standalone and aggregate calculations.
- Recalculate the aggregate. Do not assume that a quantile or other risk statistic can be added across components.
- Interpret equality correctly. No measured benefit does not contradict coherence; it may reflect aligned losses or an additive measure.
- Check the measure’s units. A mathematically computed aggregation difference may not be meaningful as a monetary incentive if the measure is not expressed in dollars.
The central lesson is that axioms shape the portfolio signal a risk measure can produce. Subadditivity rules out a measured penalty for combining losses, but it does not guarantee a strict diversification reward or determine whether the resulting risk number fits a particular management decision.
Check Your Understanding
For each problem, calculate the requested portfolio comparison and check whether the measure guarantees a nonnegative diversification benefit.
- Two mutually exclusive losses are each \$50,000 when their respective event occurs. Each event has probability 0.08, and the probability of neither event is 0.84. Find each stand-alone and aggregate 90% VaR. (Answer: Stand-alone VaRs are \$0 each; aggregate VaR is \$50,000; measured benefit is \(-\$50,000\).)
- A hypothetical loss portfolio has \(E[X]=\$7,000\) and \(E[Y]=\$11,000\). Using expected loss as the risk measure, find the aggregate measure and the diversification benefit. (Answer: Aggregate measure is \$18,000; benefit is \$0.)
- Independent losses satisfy \(X\sim N(5,3^2)\) and \(Y\sim N(9,4^2)\), with amounts in thousands of dollars. At 90% confidence, use \(z_{0.90}=1.2816\) and \(\phi(z_{0.90})/(1-0.90)=1.7550\). Find the TVaR diversification benefit. (Answer: Stand-alone sum is \(26.285\) thousand dollars; aggregate TVaR is \(22.775\) thousand dollars; benefit is \(3.510\) thousand dollars, or \$3,510.00.)
- Let \(Z\) be Bernoulli with probability 0.5 of 1. Set \(X=6Z\) and \(Y=6Z\), in dollars. Find the stand-alone variance sum, aggregate variance, and penalty under variance. (Answer: Stand-alone variance sum is 18 dollars squared; aggregate variance is 36 dollars squared; penalty is 18 dollars squared.)
- A subadditive risk measure assigns \(\rho(X)=\$24,000\), \(\rho(Y)=\$31,000\), and \(\rho(X+Y)=\$49,000\). Calculate its diversification benefit and state what the result shows. (Answer: Benefit is \$6,000; the combined portfolio has lower measured risk than the stand-alone sum, consistent with subadditivity.)