Let's take the usual definition of a spectral risk measure.

If we look at the integral we see that spectral risk measures have the property that the risk measure of a random variable $X$ can be represented by a combination of the quantiles of $X$.

Since the quantile function is rather friendly one gets that every spectral risk measure is also a coherent risk measure.

Examples are the expected value and the expected shortfall (CVaR). In those cases, the spectral representation yields a very convenient way to approximate the measure by simply weighing the quantiles of our dataset. That yields the following questions:

Are there any other known measures that have a spectral representation? If we relax the assumptions on the spectrum $\phi$, can we obtain (approximative sequences of) other (possibly non-coherent) risk measures?

EDIT: In reaction to the comment by @Joshua Ulrich I want to provide an example of what I want to achieve and some more details.

  • Example: The Conditional Value at Risk. We have the following formula: $\text{CVaR}_\alpha(X) = -\frac{1}{\alpha}\int_0^{\alpha}F^{-1}_X(p)dp$. From sample $X_i$, $i=1,\ldots,N$, we can calculate the CVaR by taking the order statistics that are in the $\alpha$-tail of the sample, average, and divide by $\alpha$. We can see that this is measure has a spectral representation with $\phi(p) = \frac{1}{\alpha}$ for $p \in [0,\alpha]$ and $\phi(p) = 0$ for $p \in (\alpha, 1]$. So its easy to check: The CVaR is a spectral measure.

Obviously, the "order statistics + weighted average" procedure does not only work for the CVaR, it works for all spectral measures: From the definition of spectral measure we see that, after discretizing the integral, we have an approximation of the measure that is a linear combination of quantiles which is very easy to compute.

In fact its so easy that I would like to compute as many risk measures as possible this way (very easy if you do monte carlo or scenarios for example). For the computation only, I dont need all the assumptions about $\phi$ so lets forget about them for a moment and see what else we can calculate this way.

  • $\begingroup$ This is a bit broad, and could lead to list-like answers. Could you provide more details on what you're actually trying to accomplish? $\endgroup$ – Joshua Ulrich Oct 20 '13 at 12:32
  • $\begingroup$ Well, I am not sure I understand what you want but what is clear is that $CVAR_{\alpha}$ works as $\Phi(p)=\frac{1}{\alpha} \cdot 1_{p \in [0,\alpha]}$ which is a density of probability. So you can build any other risk measure chosing for $\Phi$ any density on the compact set $[0,1]$, and by scaling on a compact set. I would try a "hat function" first, then sinusoids... $\endgroup$ – statquant Oct 21 '13 at 12:20
  • 1
    $\begingroup$ Quite a few risk measures seem unlikely to have a representation like this, since their units differ. For example, annualized volatility. Some risk measures, such as scenario playbacks of the '97 Asian crisis, will trivially have a spectral representation but not in any computationally useful way. $\endgroup$ – Brian B Oct 21 '13 at 13:38
  • $\begingroup$ @statquant Well, the aim is not to try a different spectrum $\phi$ but rather to identify the spectra of other well known risk measures. Brian B: I dont understand your comment about units. The units are irrelevant for a risk measure as far es the spectral property is concerned, right? $\endgroup$ – vanguard2k Oct 23 '13 at 6:47

I believe that Prospect Theory (as defined by Kahneman, Amos, and Tversky) implicitly makes use spectral risk measures. Though I am not able to find any literature linking the two, I think there is clear link between the intuitions regarding loss aversion. The key difference is that spectral risk measures are normative; we assume that the utility function is known. Prospect Theory, on the other hand, is inherently descriptive (i.e., reflects observed behaviors). Also, I am aware that spectral risk measures are extended to portfolio risk, while Prospect Theoretic measures deal with generic utility.

The value function that passes through the reference point is s-shaped and asymmetrical. The value function is steeper for losses than gains indicating that losses outweigh gains. source: Wikipedia. Prospect Theory

Again, while I haven't seen any literature on the topic, it would be interesting if someone were to show that Prospect Theoretic risk measures (which are typified by Exhibit A) meet the coherence standards for a spectral risk measure given by:

${\displaystyle \rho :{\mathcal {L}}\to \mathbb {R} }$ satisfies:

  • Positive Homogeneity: for every portfolio X and positive value ${\displaystyle \lambda >0} \lambda >0$, ${\displaystyle \rho (\lambda > X)=\lambda \rho (X)}$;
  • Translation-Invariance: for every portfolio X and $\alpha \in \mathbb {R}$, ${\displaystyle \rho (X+a)=\rho (X)-a}$;
  • Monotonicity: for all portfolios X and Y such that ${\displaystyle X\geq Y}$ , ${\displaystyle \rho (X)\leq \rho (Y)}$;
  • Sub-additivity: for all portfolios X and Y, ${\displaystyle \rho (X+Y)\leq \rho (X)+\rho (Y)}$;
  • Law-Invariance: for all portfolios X and Y with cumulative distribution functions ${\displaystyle F_{X}}$ and ${\displaystyle > F_{Y}}$ respectively, if ${\displaystyle F_{X}=F_{Y}}$ then ${\displaystyle \rho (X)=\rho (Y)}$;
  • Comonotonic Additivity: for every comonotonic random variables X and Y, ${\displaystyle \rho (X+Y)=\rho (X)+\rho (Y)}$. Note that X and Y are comonotonic if for every ${\displaystyle \omega _{1},\omega > _{2}\in \Omega :\;(X(\omega _{2})-X(\omega _{1}))(Y(\omega _{2})-Y(\omega _{1}))\geq 0}$
  • $\begingroup$ Interesting thought but thats kindof the sad story about prospect theory. It does not fulfill the necessary requirements that we have in our models. Clearly, the function above is not subadditive. Take X and -X as an example. $\endgroup$ – vanguard2k Mar 2 '18 at 12:26
  • $\begingroup$ @vanguard2k I think the function in the illustration is for a single asset, so it really doesn't say anything about how things behave in a portfolio. Regardless, my point was that if you merge the intuitions of behavioral finance into a coherent risk measure, you get something that looks a lot like a spectral risk measure. $\endgroup$ – David Addison Mar 3 '18 at 18:16

As I know, I think that spectral risk measure is a new kind of measure developed from the CVaR (weighted average value of VaR) and in the framework of coherent risk measures.

If you can prove that a risk measure is coherent then you can add any types of weighted function $\phi$ to make it a spectral risk measure. The underlying idea is that the sum of any number of coherent measures is also a coherent risk measure.

  • $\begingroup$ Utility functions just make them reflect attitude to risk of investors and sound more like behavioural finance. $\endgroup$ – David Nguyen Apr 26 at 9:42

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.