Loss-Based Risk Functionals and Sign Conventions

Risk Functionals Begin With a Convention

A risk measure is a rule that assigns a numerical risk amount to a random outcome. Before interpreting that number, specify what the random variable represents. In a loss convention, larger values are worse. In a profit-and-loss convention, larger values are better, so adverse outcomes lie farther to the left. The same portfolio can therefore have different numerical descriptions depending on the chosen sign convention.

A functional also needs a domain: the set of random variables to which the rule can be applied. This matters because some common risk calculations are finite only under suitable assumptions. For example, a bounded loss always has a finite expected value, while an unbounded loss can have a finite quantile and an infinite expected tail loss.

A standard convenient domain is \(L^\infty\), the set of random variables that are bounded except on events of probability zero. Random variables that are equal with probability one are treated as the same element of this domain. If \(|X|\leq M\) with probability one for some finite \(M\), then \(X\) has a finite expectation and every fixed probability quantile is finite. This makes \(L^\infty\) a clean setting for defining finite-valued risk functionals.

An unbounded domain can also be used, but the functional and its finiteness requirements must be stated. For instance, the expectation-based tail measure defined below is finite for an integrable loss, meaning \(E[|X|]<\infty\). It need not be finite for every almost surely finite loss. A risk calculation that produces infinity is not a finite real-valued functional on that particular input.

Loss and Profit-and-Loss Signs

Let \(X\) denote a loss, with positive values indicating payments or other adverse outcomes. Let \(Z=-X\) denote the corresponding profit-and-loss variable, with positive values indicating gains and negative values indicating losses. Given a loss-based functional \(\rho_L\), the corresponding functional under the profit-and-loss convention is defined by applying the loss functional to the negated variable:

This definition preserves the economic interpretation of the risk amount: the risk associated with a profit-and-loss variable is assessed by converting it to a loss first. It does not mean that every numerical result can be changed to a profit-and-loss result just by attaching a minus sign. In particular, quantiles depend on which side of the distribution is being measured.

The reversal can be justified directly. Suppose a loss-based functional has the property that \(A\leq B\) implies \(\rho_L(A)\leq\rho_L(B)\). If \(Z\leq W\), then \(-W\leq -Z\). Applying that loss-based property to the converted variables gives the reversed comparison below. This is an implication conditional on the stated loss-based property; defining a functional does not by itself guarantee that the property holds.

Thus, in a loss convention, the inequality is written so that a larger loss is assigned at least as much risk. In a profit-and-loss convention, a smaller profit-and-loss outcome is assigned at least as much risk. These are two expressions of the same direction of adverse movement, not contradictory rules.

Quantiles and Tail Amounts Under Each Convention

For a real-valued loss \(X\), let \(F_X(x)=P(X\leq x)\). For a probability level \(0<\alpha<1\), define the lower quantile by \(q_\alpha(X)=\inf\{x:F_X(x)\geq\alpha\}\). Under the loss convention, \(\operatorname{VaR}_\alpha(X)=q_\alpha(X)\). One tail-average convention, for an integrable loss, is

This quantile-integral definition includes the appropriate fractional contribution at a boundary quantile when the loss distribution has an atom. It describes an average of upper-tail loss amounts. For an unbounded loss, integrability is a sufficient condition for this tail average to be finite when \(0<\alpha<1\).

For a profit-and-loss variable \(Z\), lower outcomes are adverse. The corresponding tail amount is therefore obtained by applying the loss definition to \(-Z\): \(\operatorname{TVaR}_\alpha(-Z)\). Do not substitute \(Z\) directly into an upper-tail loss formula. Also, do not assume that negating a quantile at level \(\alpha\) gives the quantile of the negated variable at that same level. The quantile levels and tail directions must be handled consistently.

Why Domain and Finiteness Must Be Stated

A finite risk amount is not automatic just because a loss is a real-valued random variable. The following hypothetical severity model illustrates the distinction between a finite quantile and a finite tail average. Suppose a loss \(X\) has a Pareto survival function

This is an unbounded loss distribution with threshold \(100\) and shape parameter \(0.8\). Its distribution function for \(x\geq100\) is \(F_X(x)=1-(100/x)^{0.8}\). For a probability \(p\) between zero and one, set \(F_X(x)=p\) and solve for the quantile:

At \(p=0.95\), the quantile is \(100(0.05)^{-1.25}=100(20)^{1.25}\), approximately \(4{,}229.49\). The 95% VaR is finite. But the mean loss is not finite: for a nonnegative loss, the mean equals the integral of its survival function, and the integral of \(x^{-0.8}\) over an unbounded interval diverges. The tail-average measure at this level is also infinite. For example, the quantile-integral definition integrates \(100(1-u)^{-1.25}\) from \(u=0.95\) to \(1\), and that integral diverges at \(1\).

Consequently, this loss is not in \(L^1\), and an expectation-based tail functional is not finite on it. A finite VaR does not establish that every risk functional is finite. This is why a definition should identify both the functional’s domain and any assumptions needed to keep its value finite.

Common Sign and Domain Errors

• Switching meanings of a variable mid-calculation. State whether positive values are losses or gains. If the variable is profit-and-loss, convert it to a loss before applying a loss-based functional.
• Keeping the same inequality direction after negation. From \(Z\leq W\), the correct converted order is \(-W\leq-Z\), not \(-Z\leq-W\).
• Using the wrong tail. The upper tail is adverse for losses; the lower tail is adverse for profit-and-loss variables. Applying an upper-tail formula directly to profits and losses targets the wrong outcomes.
• Negating a quantile without checking its level. Quantiles of a negated variable concern the opposite tail. Work with the transformed variable and the stated quantile definition rather than assuming the same percentile level still applies.
• Assuming every risk amount is finite. A finite VaR does not guarantee a finite expectation or finite TVaR. Check the domain and the relevant integrability assumptions.