from mrbadguy

R is a coherent risk measure if1) if X>=0 : R(X)<=02) R(X1+X2)<=R(X1)+R(X2)3) for small k>=0 : R(kX)=k R(X)4) for a constant A : R(A+X)=R(X)-A1) if you have only positive returns, your risk capital should be non positive2) if you allocate to 2 subunits, you have enough for the whole desk3) if you multiply your portfolio, you should multiply your risk capital (k small for liquidity problems)4) if you add cash, you should take that off your risk capital

Hi Could you elaborate a bit more on 1)Why is the risk capital negative?Tronder

if you can only make money then you shouldn't need to allocate risk capital against losses

Coherent risk measures have all the right properties, the main one of which is subadditivity. VaR is not subadditive. Only with subadditivity can you measure the benefits of diversification of adding new portfolios or business lines. Furthermore, with a coherent risk measure, you can allocate capital in a fair way to the various businesses. This gives you a good way of finding the 'risk' number in the performance measures (Sharpe Ratios, etc.) that are quite popular.I'm going to try to attach one of the chapters from a course that I occasionally give on risk management, but I can't find the attacher thingy... going to clean my glasses ....Oh, poo, are attachments not allowed in this forum? Oh, yes, now I see "HTML code is not permitted." Oh well.

Call me obstinant, but I'm going to put my tex source code here. It is mostly words, so it should be ok. And $ just means maths begins/ends here. If you really think this is scintilating stuff and really can't read it, email me for a pdf.\section{VaR cannot be used for calculating diversification}If $f$ is a risk measure, the diversification benefit of aggregating portfolio's $A$ and $B$ is defined to be \begin{equation}f(A) + f(B) - f(A+B)\end{equation}When using full revaluation VaR as the methodology for computing a risk measure, it's quite possible to get negative diversification. Pathological examples are possible, but the following example is not absurd:Suppose one has a portfolio that is made up by a Trader A and Trader B. Trader A has a portfolio consisting of a put that is far out of the money, and has one day to expiry. Trader B has a portfolio that consists of a call that is also far out of the money, and also has one day to expiry. Using any historical VaR approach, say we find that each option has a probability of 4\% of ending up in the money.Trader A and B each have a portfolio that has a 96\% chance of not losing any money, so each has a 95\% VaR of zero.\footnote{To be precise, their VaR is actually a very small negative number! The average of their $V_i$'s is negative, the $5^{th}$\% is zero.} However, the combined portfolio has only a 92\% chance of not losing any money, so its VaR is non-trivial. Therefore we have a case where the risk of the combined portfolio is greater than the risks associated with the individual portfolios, i.e. negative diversification benefit if VaR is used to measure the diversification benefit. This example appears in \cite{ArtznerRiskMag}.What is so awkward about the lack of sub-additivity is the fact that this can give rise to regulatory arbitrage or to the break-down of global risk management within one single firm. This is also a serious concern for regulators. If regulation allows the capital requirement of a firm to be calculated as the sum of the requirements of its subsidiaries and if the requirements are based on VaR, the firm could create artificial subsidiaries in order to save regulatory capital.\section{Risk measures and coherence}This example introduces the concept of a ``Coherent Risk Measure''. If $f(A+B) \leq f(A) + f(B)$, where $A$ and $B$ denote portfolios, then $f$ is said to be coherent \cite{ArtznerRiskMag}, \cite{Artzner}. In fact a coherent risk measure needs to satisfy five properties \cite{Artzner}, as follows:\begin{itemize}\itemtranslation invariance: $f(A + \alpha r) = f(A) - r$, where $r$ is a reference risk free investment. (As David Heath has explained to me, this condition is simply there to ensure that the risk measure and the p\&l measure is in the same numeraire, namely, currency.)\itemSubadditivity: $f(A+B) \leq f(A) + f(B)$\itemPositive homogeneity: for all $\lambda \geq 0$, $f(\lambda A) = \lambda f(A)$.\itemMonotoneity: if $A \leq B$ then $f(A) \leq f(B)$.\itemRelevance: if $A \neq 0$ then $f(A) > 0$.\end{itemize}The property we have focused on means `a merger does not create extra risk', and is a natural requirement \cite{Artzner}.In other words the risk measure $f$ of a portfolio consisting of sub-portfolios $A$ and $B$ would always be less than or equal to the sum of the risk measure of portfolio $A$ with the risk measure of portfolio $B$. The example above shows that full revaluation VaR is not coherent. It also means that as a conservative measure of risk, one can simply add the risks calculated for the various sub-portfolios, if the measure is coherent.The earlier example is not a purely theoretical example. In practice, even on large and diverse portfolios, using VaR to calculate the diversification benefit does indeed occasionally lead to the case where this diversification is negative.There is thus a need for practical and intuitive coherent risk measures. The basic example is that in the place of a VaR calculation we use a concept known as Expected Tail Loss (ETL) or Expected Shortfall (ES). It is easiest to understand in the setting of a historical-type VaR calculation, let us say 95\% VaR. It would entail instead of taking the $5^{th}$ percentile of the p\&l's to yield a VaR number, take the average of the p\&l's up to the $5^{th}$ percentile to yield an ES number.Looking at the 5th percentile we end up with a VaR number which basically represents the best outcome of a set of bad outcomes on a bad day. Using ES we look at an average bad outcome on a bad day. This ES number turns out to be a coherent risk measure \cite{Artzner}, \cite{Eber}, \cite{AcerbiTasche}, and therefore guarantees that the diversification is always positive. As stated in the abstract of \cite{AcerbiTasche}, ``Expected Shortfall (ES) in several variants has been proposed as remedy for the deficiencies of Value-at-Risk (VaR) which in general is not a coherent risk measure''.A readable account of these and related issues is \cite{ANS}.One should report both VaR and ES, but use only ES to calculate and report diversification. Note that because standard deviation is sub-additive the standard RiskMetrics simplification is coherent. The usual RiskMetrics VaR is also subadditive (and hence coherent), but this is a mathematical exercise for masochists - it is not easy at all. \section{Measuring diversification}The diversification benefit of portfolio $P_0$ is equal to $$f \left ( {\sum_{i=1}^n P_i} \right ) + f(P_0) - f \left ( {\sum_{i=0}^n P_i} \right )$$where $f$ denotes ES and $P_1, \: P_2, \: \ldots, \: P_n$ are the (original) portfolios against which the diversification is measured.\section{Coherent capital allocation}The intention is to allocate capital costs in a coherent manner. This sounds like quite an otherworldly exercise, but one can make the task quite concrete and ask: of my risk number (such as ES), how much (as a percentage, say) is due to each of my traders? Then, given my capital adequacy charges (which may or may not be calculated via an approved internal model) I can allocate as a cost the charges in those proportions to each of those desks.Each desk can break down their own charges amongst their dealers, and management can decide where the greatest risk management focus needs to lie.\cite{Denault} has developed a method of allocating the risk capital costs to the various sub-portfolios in a fair manner, yielding for each portfolio, a risk appraisal that takes diversification into account. We wish to thank Freddie Delbaen, who contributed significantly to that paper, for clarifying certain issues.The approach of \cite{Denault} is axiomatic, starting from a risk measure which is coherent in the above sense. We may specialise the results of \cite{Denault} to the case of the coherent Expected Shortfall risk measure in which case his results become quite concrete.An allocation method for risk capital is then said to be coherent if \begin{itemize}\itemThe risk capital is fully allocated to the portfolios, in particular, each portfolio can be assigned a percentage of the total risk capital.\itemThere is `no undercut': no portfolio's allocation is higher than if they stood alone. Similarly for any coalition of portfolios and coalition of fractional portfolios.\item`Symmetry': a portfolio's allocation depends only on its contribution to risk within the firm, and nothing else.\item`Riskless allocation': a portfolio that increases its cash position will see its allocated capital decrease by the same amount.\end{itemize}All of these requirements have precise mathematical formulations.A coherent allocation is to be understood as one that is fair and credible. One should not be surprised to be told that this is a game theoretic problem where the portfolios are players, looking for their own optimal strategy. \cite{Denault} applies some results from game theory to show that the so-called Aumann-Shapley value from game theory is an appropriate allocation (it is a Nash equilibrium). Further, some results (fairly easy to derive in this special case) from \cite{Tasche} on the differentiability of Expected Shortfall show that the Aumann-Shapley value is given by \begin{equation}K_i = - \mathbb E [X_i | X \leq q_\alpha ] \end{equation}where $X_i$ denotes the (random vector of) p\&l's of the $i^{th}$ portfolio, $X = \sum_j X_j$ is the vector of p\&l's of the company, and $q_\alpha$ is the $\alpha$ percentile of $X$.Hence, as a percentage of total capital, the capital cost for the $i^{th}$ portfolio is \begin{equation}\frac{\mathbb E [X_i | X \leq q_\alpha ]}{\mathbb E [X | X \leq q_\alpha]}\end{equation}In the context of any historical or Monte Carlo type VaR model, this fraction is easy to calculate: \begin{itemize}\itemThe denominator is the average of the $1 - \alpha$\% worst p\&l's of the entire bank,\itemThe numerator is the average of the p\&l's that correspond to the same experiments as in the denominator.\end{itemize}\section{Greek Attribution}The same procedure can be followed to attribute the risk to the various Greeks. Suppose, for simplicity, that we have a single equity option and we wish to decompose the risks. Assume that we are using a historical-type or Monte Carlo method for calculating our VaR or ES. Then we can consider the p\&l's generated by the various historical or Monte Carlo experiments. Of course, our risk measure is calculated by looking at the tail of this distribution of p\&l's.By considering a first order (delta-gamma-rho-vega-theta) Taylor series expansion as follows:\begin{eqnarray*} dV& \simeq & \delta \Delta S + \half \gamma \Delta S^2 + \rho \Delta{r} + \Vega \Delta \sigma + \theta \Delta t\end{eqnarray*}we can attribute the p\&l in each experiment as \begin{eqnarray*} dV_i& = & \Delta (S_i-S) + \half \Gamma (S_i-S)^2 + \rho (r_i-r) + \Vega (\sigma_i-\sigma) + \theta \delta t + \epsilon_i\end{eqnarray*}where $S$ is the original spot, $S_i$ is the $i^{th}$ spot experiment, etc. The p\&l's due to delta are the $\Delta(S_i-S)$, and we can attribute to delta the proportion of the ES of the entire position. The same follows for the remaining greeks. One does not include higher order (mixed) partial derivatives in this expansion, because such effects will be implicit in the historical experiments one is creating (and should be implicit in the Monte Carlo, if the generator is built sufficiently well).One things need to be checked for: namely, that the $\epsilon_i$ are not material. Of course, the method is attributing a percentage to this error term, which should not be more than a couple of percent. After all, the error term is a measure of how well the Taylor series expansion is fitting the actual p\&l. As expected, for more complicated products, these errors can be more material, and the method should not be used.\bibitem{ArtznerRiskMag}Philippe Artzner, Freddy Delbaen, Jean-Marc Eber, and David Heath.\newblock Thinking coherently.\newblock {\em Risk}, 10:68--71, November 1997.\bibitem{Artzner}Philippe Artzner, Freddy Delbaen, Jean-Marc Eber, and David Heath.\newblock Coherent measures of risk.\newblock {\em Math. Finance}, 9(3):203--228, 1999.\newblock www.math.ethz.ch/{$\tilde{}$}delbaen.\b ... sche}Carlo Acerbi and Dirk Tasche.\newblock On the coherence of expected shortfall.\newblock www-m1.mathematik.tu-muenchen.de/m4/Papers/Tasche/shortfall.pdf.\bibitem{ANS}Carlo Acerbi, Claudio Nordio, and Carlo Sirtori.\newblock Expected shortfall as a tool for financial risk management.\newblock www.gloriamundi.org/var/wps.html, 2001.\bibitem{Denault}Michel Denault.\newblock Coherent allocation of risk capital.\newblock {\em Journal of Risk}, 4(1):1--34, 2001.\bibitem{Tasche}D.~Tasche.\newblock Risk contributions and performance measurement.\newblock {\em Technische {Universit\"{a}t} {M\"{u}nchen}}, TUM-M9909, July 1999.

Graeme, we used the Equation Editor to make the gifs of your LaTeX. It was rather on the long side so we had to break it up into smaller pieces...that's why the references are not coherent

A digestible version of the above document is now here.