Diversification in Portfolio Contingency Loads

Why Portfolio Risk Depends on More Than Segment Risk

Two underwriting segments can each have substantial uncertainty while the combined portfolio is less variable than the sum of their standalone risks. The reason is that segment losses may not move together. When losses in one segment tend to be unusually high at the same time as losses in the other, portfolio risk increases; when their movements are less aligned, some fluctuations offset one another.

This matters when a contingency provision is selected using a measure of risk above expected losses. Adding the separate segment provisions may be conservative, but it ignores how the segments interact. A portfolio calculation uses both segments’ variances and their dependence. It does not, by itself, determine how the total provision should be assigned back to each segment.

Definition: Let $L_1$ and $L_2$ be the aggregate losses for two underwriting segments over the same accounting or pricing period and on a consistent basis. Write $\mu_j=E[L_j]$ and $\sigma_j=\sqrt{\operatorname{Var}(L_j)}$ for segment $j$. If both standard deviations are positive, their correlation is $\rho=\operatorname{Cov}(L_1,L_2)/(\sigma_1\sigma_2)$. The portfolio aggregate loss is $L=L_1+L_2$.

Aggregating Variance and Standard Deviation

For finite-variance losses, the variance of the portfolio is the sum of the segment variances and twice their covariance. Expressing covariance through correlation gives a convenient calculation formula:

\operatorname{Var}(L) =\sigma_1^2+\sigma_2^2+2\operatorname{Cov}(L_1,L_2) =\sigma_1^2+\sigma_2^2+2\rho\sigma_1\sigma_2.

The portfolio standard deviation is the square root of this variance. Correlation is between $-1$ and $1$; a higher value means stronger positive linear association, not necessarily a particular causal relationship. Correlation alone does not say that losses are normally distributed.

Why the Covariance Term Appears

The variance formula follows by centering each segment loss at its expectation. Since $E[L]=\mu_1+\mu_2$, the deviation of total loss from its expectation is $(L_1-\mu_1)+(L_2-\mu_2)$. Squaring this expression produces two squared deviations and a cross-product:

\begin{aligned} \operatorname{Var}(L) &=E\left[\left((L_1-\mu_1)+(L_2-\mu_2)\right)^2\right]\\ &=E[(L_1-\mu_1)^2]+E[(L_2-\mu_2)^2]\\ &\quad+2E[(L_1-\mu_1)(L_2-\mu_2)]\\ &=\sigma_1^2+\sigma_2^2+2\operatorname{Cov}(L_1,L_2). \end{aligned}

The last expectation is the covariance by definition. Replacing it with $\rho\sigma_1\sigma_2$ gives the correlation form. Thus, the covariance term is not an optional adjustment: it is the contribution of joint movement in the two segment losses.

Key result: If a contingency rule sets the load equal to $k$ times standard deviation, where $k\geq 0$ is a selected factor, then the portfolio load is $k\sqrt{\sigma_1^2+\sigma_2^2+2\rho\sigma_1\sigma_2}$. This is generally not equal to the sum of the standalone loads $k\sigma_1+k\sigma_2$.

A standard-deviation rule is a chosen risk measure; the variance identity does not require that losses have any particular distribution. A probability-based interpretation needs additional assumptions. For example, if $L$ is normally distributed, then its one-sided $p$-quantile is $\mu_1+\mu_2+z_p\sigma_L$, where $z_p$ is the standard normal quantile and $\sigma_L$ is the portfolio standard deviation. Without normality (or another specified distribution), this quantile formula does not follow from the means and variances alone.

Worked Example: A Standard-Deviation Load for Two Segments

A hypothetical insurer writes two segments over the same year. Segment A has expected aggregate losses of $3.2 million and standard deviation $0.6 million. Segment B has expected aggregate losses of $2.5 million and standard deviation $0.8 million. Their loss correlation is $0.25$. The insurer uses a contingency rule of $0.75$ times standard deviation. Calculate the portfolio load and compare it with the sum of the standalone loads.

Method and assumptions. Use the variance aggregation formula. The standard-deviation rule is applied to the loss uncertainty above expected losses; no probability distribution is needed for this rule.

Step 1: Calculate portfolio expected losses. Linearity of expectation gives $3.2+2.5=5.7$ million dollars. Correlation does not change this expected total.

Step 2: Calculate portfolio variance. In squared millions of dollars, $\sigma_L^2=0.6^2+0.8^2+2(0.25)(0.6)(0.8)=0.36+0.64+0.24=1.24$.

Step 3: Calculate portfolio standard deviation and load. The standard deviation is $\sqrt{1.24}=1.11355$ million dollars. The portfolio load is $0.75(1.11355)=0.83516$ million dollars.

Step 4: Compare standalone loads. The separate loads are $0.75(0.6)=0.45$ million and $0.75(0.8)=0.60$ million, totaling $1.05$ million. The portfolio-based load is lower by $1.05-0.83516=0.21484$ million dollars.

Check. The correlation is positive, so the portfolio standard deviation is greater than the independent-risk value $\sqrt{0.36+0.64}=1.0$ million. It remains below the sum of segment standard deviations, $1.4$ million. Both comparisons are consistent with the calculation. The common mistake avoided here is adding standard deviations before applying the load factor.

How Correlation Changes the Result

Relationship between segment losses	Effect on portfolio variance	Interpretation
Positive correlation	The covariance term is positive.	Losses tend to move in the same direction, increasing aggregate risk relative to independence.
Zero correlation	The covariance term is zero.	Variances add. Zero correlation does not, in general, mean statistical independence.
Negative correlation	The covariance term is negative.	Opposing movements can reduce aggregate risk, but the resulting variance cannot be negative.

For perfectly positive correlation, $\rho=1$, the portfolio standard deviation is $\sigma_1+\sigma_2$. For perfectly negative correlation, $\rho=-1$, it is $|\sigma_1-\sigma_2|$. These boundary cases follow by taking the square root of $(\sigma_1+\sigma_2)^2$ and $(\sigma_1-\sigma_2)^2$, respectively. They also show why diversification reduces risk only to the extent that segment losses are not perfectly positively aligned.

Worked Example: Negative Correlation and a Lower Load

A hypothetical insurer has two small specialty segments, each with expected annual aggregate losses of $1.1 million and a standard deviation of $0.5 million. Assume the segments’ aggregate losses have correlation $-0.5$. Under a one-standard-deviation contingency rule, calculate the portfolio load and the reduction from the sum of standalone loads.

Method and assumptions. Apply the variance formula to the two segment losses. The problem specifies a standard-deviation load, not a normal quantile.

Step 1: Find the expected total. The total expected loss is $1.1+1.1=2.2$ million dollars.

Step 2: Find covariance and portfolio variance. Covariance is $(-0.5)(0.5)(0.5)=-0.125$ million-squared dollars. Therefore, $\sigma_L^2=0.25+0.25+2(-0.125)=0.25$ million-squared dollars.

Step 3: Find the portfolio load. The portfolio standard deviation is $\sqrt{0.25}=0.5$ million dollars. At one times standard deviation, the portfolio load is $0.5 million.

Step 4: Compare with separate loads. Each standalone load is $0.5 million, so their sum is $1.0 million. The portfolio load is lower by $1.0-0.5=$0.5$ million.

Check. The result equals the standard deviation of either segment because the equal-sized risks partially offset at this correlation; it is not zero. The common mistake avoided is treating negative correlation as eliminating all risk.

A Probability Target Requires a Distribution

A contingency provision may instead be described as the amount needed to reach a target probability of covering aggregate losses. Such a statement is stronger than specifying a standard-deviation multiple: it requires a distributional model or enough information to determine the relevant quantile. Under a joint normal model for segment aggregate losses, their sum is normal, and its mean and variance determine the portfolio quantile. A normal approximation may be useful, but it must be stated rather than inferred from correlation.

Worked Example: A Normal-Model 95% Loss Threshold

For a hypothetical commercial insurance portfolio, assume the annual aggregate losses in segments A and B are jointly normal. Their means are $2.4 million and $1.6 million, their standard deviations are $0.3 million and $0.4 million, and their correlation is $0.25$. Estimate the 95th percentile of total losses and compare the contingency amount above expected losses with the sum of the two separate 95% normal contingency amounts.

Method and assumptions. Joint normality makes the sum normal. Use the one-sided standard normal quantile $z_{0.95}=1.644854$. The requested contingency amount is the 95th percentile less expected losses.

Step 1: Aggregate the means. The portfolio mean is $2.4+1.6=4.0$ million dollars.

Step 2: Aggregate the variances. The covariance is $(0.25)(0.3)(0.4)=0.03$ million-squared dollars. Thus, $\sigma_L^2=0.3^2+0.4^2+2(0.03)=0.09+0.16+0.06=0.31$ million-squared dollars, and $\sigma_L=\sqrt{0.31}=0.556776$ million dollars.

Step 3: Calculate the portfolio contingency amount. The amount above the mean is $1.644854(0.556776)=0.91582$ million dollars. The 95th-percentile total loss is $4.0+0.91582=4.91582$ million dollars.

Step 4: Compare separate amounts. The separate 95% amounts above each segment mean sum to $1.644854(0.3)+1.644854(0.4)=1.644854(0.7)=1.15140$ million dollars. The portfolio amount is lower by $1.15140-0.91582=0.23558$ million dollars.

Check. The portfolio standard deviation $0.556776$ million is less than the sum $0.7$ million, but greater than the independent value $\sqrt{0.09+0.16}=0.5$ million because correlation is positive. The common mistake avoided is using the normal quantile without stating a normal-model assumption.

Common Exam Trap: Confusing Correlation with Covariance

Correlation is dimensionless; covariance has squared-loss units. The variance formula requires either $2\operatorname{Cov}(L_1,L_2)$ or $2\rho\sigma_1\sigma_2$. Substituting correlation alone as if it were covariance gives an incorrect result and inconsistent units.

For example, if $\sigma_1=0.6$ million, $\sigma_2=0.8$ million, and $\rho=0.25$, the covariance is $0.25(0.6)(0.8)=0.12$ million-squared dollars. The covariance contribution is $2(0.12)=0.24$ million-squared dollars. A wrong calculation might use $2(0.25)=0.50$ as the contribution. That is not the specified covariance contribution: it omits both segment standard deviations and does not have the required squared-loss units.

A further check is to confirm that the resulting variance is nonnegative. With valid standard deviations and a valid correlation, the formula must satisfy $\sigma_L^2\geq 0$. If it does not, recheck the units, signs, and multiplication.

Key takeaway: A portfolio contingency load depends on both segment volatility and dependence. Calculate aggregate variance first, take its square root, and then apply the stated load rule. Use a probability-based interpretation only when a suitable distributional assumption has been supplied.

Check Your Understanding

For each problem, state the risk measure or assumptions used and keep variance units consistent.

Two hypothetical segments have standard deviations of $200,000 and $500,000 and loss correlation $0.4$. Under a $0.6$-times-standard-deviation rule, calculate the portfolio load. (Answer: approximately $364,966.)
Two segments have standard deviations of $300,000 and $400,000 and zero correlation. Compare the portfolio load under a $0.8$-times-standard-deviation rule with the sum of the standalone loads. (Answer: portfolio load $400,000; standalone loads total $560,000; reduction $160,000.)
Two segments each have standard deviation $250,000 and loss correlation $-0.28$. Calculate the portfolio standard deviation and the load under a $1.2$-times-standard-deviation rule. (Answer: standard deviation $300,000; load $360,000.)
Segment A has standard deviation $400,000 and segment B has standard deviation $500,000. What correlation would produce a portfolio standard deviation of $600,000? (Answer: $-0.125$.)
Assume joint normality. Two segment aggregate losses have means of $1.0 million and $0.7 million, standard deviations of $0.2 million and $0.3 million, and correlation $0.5$. Estimate the portfolio 90th-percentile loss using $z_{0.90}=1.2816$. (Answer: approximately $2.259 million.)

What You'll Learn