Monotonicity as Ordering of Loss Distributions

Why Ordering Losses Matters

A risk measure used to assess losses should not assign a lower risk amount to a loss that is larger in every relevant scenario. This principle is called monotonicity. The central comparison is made scenario by scenario: it asks whether one loss is no greater than another, except possibly on an event of probability zero.

That condition is stronger than comparing expected losses. A smaller mean does not guarantee that one loss is smaller in every scenario, and a larger mean does not by itself establish which loss has the greater VaR. The distinction matters when applying risk measures to hypothetical portfolios, insurance losses, or other modeled exposures. The definitions below use the loss-positive convention: larger values of $X$ represent larger losses.

Definition: A loss-based risk measure $\rho$ is monotone if, for losses $X$ and $Y$, $$ P(X\leq Y)=1 \quad\Longrightarrow\quad \rho(X)\leq \rho(Y). $$ The condition $P(X\leq Y)=1$ means $X\leq Y$ almost surely (a.s.). It permits exceptions only on events of probability zero; it does not mean merely that $X\leq Y$ in most scenarios.

The losses in this definition are defined on the same probability space, so their outcomes can be compared scenario by scenario. This joint specification is a coupling of the losses. A comparison of two separate probability distributions may show that one distribution is generally shifted toward larger losses, but it does not automatically tell us how the two losses compare on a particular shared set of scenarios.

Quantiles, VaR, and TVaR

Let $F_X(x)=P(X\leq x)$ be the cumulative distribution function of a loss $X$. For $0<u<1$, define the lower generalized inverse, or $u$-quantile, by

q_X(u)=\inf\{x:F_X(x)\geq u\}.

For confidence level $\alpha$, where $0<\alpha<1$, define $\operatorname{VaR}_{\alpha}(X)=q_X(\alpha)$. In a discrete loss distribution, this definition matters: the quantile is the smallest loss amount at which the cumulative probability reaches or exceeds $\alpha$.

For this tutorial, define TVaR using the upper quantile integral. Assume losses are finite and have finite means, so the integral below is finite. This definition handles discrete distributions as well as continuous ones, including distributions with atoms.

Definition: For $0<\alpha<1$, $$ \operatorname{TVaR}_{\alpha}(X) =\frac{1}{1-\alpha}\int_{\alpha}^{1}q_X(u)\,du. $$ It is the average of the quantiles over the upper $1-\alpha$ fraction of probability. When a distribution has atoms, the integral definition accounts for the appropriate fraction of probability at a boundary; it is not always correct to use only losses strictly exceeding VaR.

VaR and TVaR are both monotone. The reason is that almost-sure ordering places an ordering on the cumulative distribution functions, which in turn places an ordering on every quantile. TVaR then averages those ordered quantiles over the same range of probability levels.

Why Quantiles Preserve the Loss Ordering

Suppose $P(X\leq Y)=1$. For any real number $t$, the event $\{Y\leq t\}$ is contained, up to probability-zero exceptions, in $\{X\leq t\}$: if $Y\leq t$ and $X\leq Y$, then $X\leq t$. Therefore,

F_X(t)=P(X\leq t)\geq P(Y\leq t)=F_Y(t).

To pass from this CDF ordering to a quantile ordering, fix $u$ between 0 and 1. If $F_Y(t)\geq u$, then $F_X(t)\geq F_Y(t)\geq u$. Thus every $t$ in the set used to define $q_Y(u)$ is also in the set used to define $q_X(u)$. Taking the infimum gives $q_X(u)\leq q_Y(u)$. This argument proves the quantile ordering rather than assuming that the losses have continuous distributions.

Start with the scenario ordering.
Assume $P(X\leq Y)=1$, with both losses defined on the same probability space.

Order the CDFs.
For every real $t$, $F_X(t)\geq F_Y(t)$, because $Y\leq t$ implies $X\leq t$ outside a probability-zero event.

Order the quantiles.
For every $0<u<1$, the set $\{t:F_Y(t)\geq u\}$ is contained in $\{t:F_X(t)\geq u\}$. Hence $q_X(u)\leq q_Y(u)$.

Apply the risk measure.
At $u=\alpha$, the quantile inequality gives $\operatorname{VaR}_{\alpha}(X)\leq\operatorname{VaR}_{\alpha}(Y)$. Integrating the inequality over $[\alpha,1]$ and dividing by $1-\alpha$ gives $\operatorname{TVaR}_{\alpha}(X)\leq\operatorname{TVaR}_{\alpha}(Y)$.

Theorem: For finite losses $X$ and $Y$ with finite means, if $P(X\leq Y)=1$, then for every $0<\alpha<1$, $$ \operatorname{VaR}_{\alpha}(X)\leq \operatorname{VaR}_{\alpha}(Y) \quad\text{and}\quad \operatorname{TVaR}_{\alpha}(X)\leq \operatorname{TVaR}_{\alpha}(Y). $$

Worked Examples

Worked Example: Ordering VaR and TVaR in a Coupled Loss Model

A hypothetical insurer models two one-period losses, $X$ and $Y$, on the same two equally specified scenarios. Scenario 1 has probability $0.4$; scenario 2 has probability $0.6$. Losses, in thousands of dollars, are $X=(10,40)$ and $Y=(15,50)$, respectively. There is no interest or mortality component. At $\alpha=0.4$, calculate each loss's mean, VaR, and TVaR, and verify the ordering.

Method and ordering check. Compare the losses in each shared scenario: $10\leq15$ in scenario 1 and $40\leq50$ in scenario 2. Therefore $P(X\leq Y)=1$, so the monotonicity result applies.

Means. For $X$, $E[X]=0.4(10)+0.6(40)=4+24=28$ thousand dollars. For $Y$, $E[Y]=0.4(15)+0.6(50)=6+30=36$ thousand dollars.

VaR. The cumulative probability for $X$ reaches $0.4$ at loss 10, so $\operatorname{VaR}_{0.4}(X)=10$. For $Y$, it reaches $0.4$ at loss 15, so $\operatorname{VaR}_{0.4}(Y)=15$.

TVaR. Above probability level $0.4$, the quantile for $X$ is 40 throughout the remaining probability interval, while the quantile for $Y$ is 50. Thus $\operatorname{TVaR}_{0.4}(X)=40$ and $\operatorname{TVaR}_{0.4}(Y)=50$, in thousands of dollars.

The results check: $28\leq36$, $10\leq15$, and $40\leq50$. The method avoids the common mistake of comparing only means: the scenario-by-scenario comparison establishes the condition needed for monotonicity.

Worked Example: Using All Quantiles in a Four-Scenario Model

A hypothetical risk manager assigns probability $0.25$ to each of four scenarios. The losses, in thousands of dollars, are $X=(0,10,30,30)$ and $Y=(5,15,35,50)$, listed in the same scenario order. There is no interest or mortality component. At $\alpha=0.5$, calculate VaR and TVaR and explain why the quantile comparison works even though $X$ and $Y$ have different distributions.

Check the coupling. In order, the scenario comparisons are $0\leq5$, $10\leq15$, $30\leq35$, and $30\leq50$. Thus $X\leq Y$ in every scenario.

Find VaR. The ordered values of $X$ are $0,10,30,30$; their cumulative probabilities are $0.25,0.50,0.75,1$. Therefore $\operatorname{VaR}_{0.5}(X)=10$. The ordered values of $Y$ are $5,15,35,50$, so $\operatorname{VaR}_{0.5}(Y)=15$.

Find TVaR. The upper half of the probability distribution corresponds to the two largest values, each with probability $0.25$. For $X$, those values are 30 and 30, giving $\operatorname{TVaR}_{0.5}(X)=(30+30)/2=30$. For $Y$, they are 35 and 50, giving $\operatorname{TVaR}_{0.5}(Y)=(35+50)/2=42.5$, in thousands of dollars.

The ordering checks: $10\leq15$ for VaR and $30\leq42.5$ for TVaR. The distributions differ, but the proof does not require them to be the same. It uses the shared scenarios to establish almost-sure ordering, then the resulting quantile ordering.

Worked Example: Why Comparing Means Is Not Enough

A hypothetical analyst compares two one-period losses, each with two equally likely outcomes. In the stated coupling, $X=(20,80)$ and $Y=(60,70)$, in thousands of dollars. There is no interest or mortality component. Calculate their means and their VaR at $\alpha=0.75$, and determine whether either loss is almost surely no greater than the other.

Means. $E[X]=0.5(20)+0.5(80)=10+40=50$ thousand dollars. $E[Y]=0.5(60)+0.5(70)=30+35=65$ thousand dollars.

VaR. For $X$, the CDF reaches $0.75$ at 80, so $\operatorname{VaR}_{0.75}(X)=80$. For $Y$, it reaches $0.75$ at 70, so $\operatorname{VaR}_{0.75}(Y)=70$, in thousands of dollars.

Check the scenario ordering. In scenario 1, $20\leq60$, but in scenario 2, $80>70$, so $X\leq Y$ does not hold almost surely. Conversely, $60>20$ in scenario 1, so $Y\leq X$ does not hold almost surely either.

The mean comparison $50<65$ does not establish almost-sure ordering; indeed, the VaR comparison goes in the opposite direction. A wrong approach would be to infer from $E[X]<E[Y]$ that $X\leq Y$ a.s. and then expect every risk measure to rank $X$ lower. That inference is invalid: a mean summarizes average loss, while almost-sure ordering compares losses in each scenario. The example avoids conflating those distinct comparisons.

Distribution Comparisons and Common Traps

A useful related fact follows from the proof: if $F_X(t)\geq F_Y(t)$ for every $t$, then $q_X(u)\leq q_Y(u)$ for every $u$, and therefore VaR and TVaR are ordered. This is a comparison of the marginal distributions, often expressed as first-order stochastic ordering. It does not require a particular coupling to prove the quantile inequalities. However, it is different from claiming that two losses in a specified model satisfy $X\leq Y$ almost surely. A marginal distribution comparison does not tell us that the losses are ordered in every paired scenario of an arbitrary coupling.

The logical direction is important. Almost-sure ordering implies the CDF and quantile orderings used above. If that scenario-by-scenario condition fails, the monotonicity theorem simply does not decide the comparison. It does not follow that VaR or TVaR must be ordered either way.

Comparison being made	What it establishes	What it does not establish
$X\leq Y$ almost surely in a specified coupling	Quantile, VaR, and TVaR of $X$ are no greater than those of $Y$	That the means are equal or that the inequality is strict
$E[X]\leq E[Y]$	A comparison of average losses, when the means exist	Almost-sure ordering or a VaR/TVaR ordering
$F_X(t)\geq F_Y(t)$ for all $t$	Quantile ordering, and hence VaR and TVaR ordering	Scenario-by-scenario ordering in an arbitrary specified coupling

Confusing “usually smaller” with almost surely smaller. Monotonicity requires the inequality to hold except on a probability-zero event, not merely in most scenarios.
Using the mean as a substitute for a loss ordering. Expected values do not determine how losses compare across scenarios or at a selected quantile.
Using the wrong quantile convention for a discrete loss. VaR is the smallest amount at which the CDF reaches $\alpha$; it is not automatically the next larger outcome.
Forgetting what monotonicity does not say. When losses are not ordered, monotonicity alone gives no ranking of their risk measures.

Key takeaway: For quantile-based risk measures, the proof of monotonicity runs from an ordering of losses to an ordering of distributions, then quantiles, then VaR and TVaR. Comparing means alone skips the essential ordering condition. This material is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus.

Check Your Understanding

For each problem, identify whether a scenario-by-scenario ordering is given or can be verified before drawing conclusions about the risk measures.

A hypothetical loss model has three scenarios with probabilities $0.2,0.5,0.3$. Losses in thousands of dollars are $X=(5,25,40)$ and $Y=(8,25,55)$ in the same scenario order. At $\alpha=0.7$, find each mean, VaR, and TVaR. (Answer: $E[X]=25.5$, $E[Y]=30.6$; VaR values are 25 and 25; TVaR values are 40 and 55, respectively, all in thousands of dollars.)
Four equally likely scenarios have losses $X=(0,10,20,50)$ and $Y=(2,12,25,55)$, in thousands of dollars. At $\alpha=0.5$, find each VaR and TVaR and state whether monotonicity applies. (Answer: $X\leq Y$ in every scenario; VaR values are 10 and 12, and TVaR values are 35 and 40, respectively, in thousands of dollars.)
Two equally likely outcomes give losses $X=(0,100)$ and $Y=(40,60)$, in thousands of dollars, under the stated coupling. Compare their means and their VaR at $\alpha=0.75$. Is either loss almost surely no greater than the other? (Answer: Both means are 50 thousand dollars; VaR values are 100 and 60 thousand dollars; neither loss is almost surely no greater than the other.)
Two marginal loss distributions are $X=0$ with probability $0.6$ and $X=30$ with probability $0.4$, and $Y=10$ with probability $0.6$ and $Y=40$ with probability $0.4$, all amounts in thousands of dollars. At $\alpha=0.6$, compare their means, VaR, and TVaR using the marginal distributions. (Answer: $E[X]=12$, $E[Y]=22$; VaR values are 0 and 10; TVaR values are 30 and 40, respectively, in thousands of dollars.)
Three equally likely scenarios have coupled losses $X=(10,50,20)$ and $Y=(15,45,30)$, in thousands of dollars. At $\alpha=0.67$, find the VaR values and determine whether either loss is almost surely no greater than the other. (Answer: The ordered values give VaR values of 50 for $X$ and 45 for $Y$, in thousands of dollars; neither loss is almost surely no greater than the other.)

Comparison being made	What it establishes	What it does not establish
\(X\leq Y\) almost surely in a specified coupling	Quantile, VaR, and TVaR of \(X\) are no greater than those of \(Y\)	That the means are equal or that the inequality is strict
\(E[X]\leq E[Y]\)	A comparison of average losses, when the means exist	Almost-sure ordering or a VaR/TVaR ordering
\(F_X(t)\geq F_Y(t)\) for all \(t\)	Quantile ordering, and hence VaR and TVaR ordering	Scenario-by-scenario ordering in an arbitrary specified coupling

What You'll Learn