In last few years Value-at-Risk (Var) is a very popular and frequently used risk measure. Risk measure VaR is used in most of the financial institutions. VaR is popular thanks to its simple interpretation and simple valuation. Valuation of VaR is a problem if we assume a few dependent risks. So VaR is estimated in a practice. In presented thesis we study theory of stochastic bounding. Using this theory we obtain bounds for VaR of sum a few dependent risks. In next part of presented thesis we show how we can generalize obtained bounds by theory of copulae. Then we show numerical algorithm, which we can use to evaluate bounds, when exact analytical evaluate isn't possible. In a final part of presented thesis we show our results on practical examples.

This dissertation consists of three essays. In the first essay, I estimate a high dimensional covariance matrix of returns for 88 individual stocks from the S&P 100 index, using daily return data for 1995-2005. This study applies the two-step estimator of the dynamic conditional correlation multivariate GARCH model, proposed by Engle (2002b) and Engle and Sheppard (2001) and applies variations of this model. This is the first study estimating variances and covariances of returns using a large number of individual stocks (e.g., Engle and Sheppard (2001) use data on various aggregate sub-indexes of stocks). This avoids errors in estimation of GARCH models with contemporaneous aggregation of stocks (e.g. Nijman and Sentana 1996; Komunjer 2001). Second, this is the first multivariate GARCH adopting a systematic general-to-specific approach to specification of lagged returns in the mean equation. Various alternatives to simple GARCH are considered in step one univariate estimation, and econometric results favour an asymmetric EGARCH extension of Engle and Sheppard’s model. In essay two, I aggregate a variance-covariance matrix of return risk (estimated using DCC-MVGARCH in essay one) to an aggregate index of return risk. This measure of risk is compared with the standard approach to measuring risk from a simple univariate GARCH model of aggregate returns. In principle the standard approach implies errors in estimation due to contemporaneous aggregation of stocks. The two measures are compared in terms of correlation and economic values: measures are not perfectly correlated, and the economic value for the improved estimate of risk as calculated here is substantial. Essay three has three parts. The major part is an empirical study of the aggregate risk return tradeoff for U.S. stocks using daily data. Recent research indicates that past risk-return studies suffer from inadequate sample size, and this suggests using daily rather than monthly data. Modeling dynamics/lags is critical in daily models, and apparently this is the first such study to model lags correctly using a general to specific approach. This is also the first risk return study to apply Wu tests for possible problems of endogeneity/measurement error for the risk variable. Results indicate a statistically significant positive relation between expected returns and risk, as is predicted by capital asset pricing models. Development of the Wu test leads naturally into a model relating aggregate risk of returns to economic variables from the risk return study. This is the first such model to include lags in variables based on a general to specific methodology and to include covariances of such variables. I also derive coefficient links between such models and risk-return models, so in theory these models are more closely related than has been realized in past literature. Empirical results for the daily model are consistent with theory and indicate that the economic and financial variables explain a substantial part of variation in daily risk of returns. The first section of this essay also investigates at a theoretical and empirical level several alternative index number approaches for aggregating multivariate risk over stocks. The empirical results indicate that these indexes are highly correlated for this data set, so only the simplest indexes are used in the remainder of the essay.

This dissertation consists of three essays. In the first essay, I estimate a high dimensional covariance matrix of returns for 88 individual stocks from the S&P 100 index, using daily return data for 1995-2005. This study applies the two-step estimator of the dynamic conditional correlation multivariate GARCH model, proposed by Engle (2002b) and Engle and Sheppard (2001) and applies variations of this model. This is the first study estimating variances and covariances of returns using a large number of individual stocks (e.g., Engle and Sheppard (2001) use data on various aggregate sub-indexes of stocks). This avoids errors in estimation of GARCH models with contemporaneous aggregation of stocks (e.g. Nijman and Sentana 1996; Komunjer 2001). Second, this is the first multivariate GARCH adopting a systematic general-to-specific approach to specification of lagged returns in the mean equation. Various alternatives to simple GARCH are considered in step one univariate estimation, and econometric results favour an asymmetric EGARCH extension of Engle and Sheppard’s model. In essay two, I aggregate a variance-covariance matrix of return risk (estimated using DCC-MVGARCH in essay one) to an aggregate index of return risk. This measure of risk is compared with the standard approach to measuring risk from a simple univariate GARCH model of aggregate returns. In principle the standard approach implies errors in estimation due to contemporaneous aggregation of stocks. The two measures are compared in terms of correlation and economic values: measures are not perfectly correlated, and the economic value for the improved estimate of risk as calculated here is substantial. Essay three has three parts. The major part is an empirical study of the aggregate risk return tradeoff for U.S. stocks using daily data. Recent research indicates that past risk-return studies suffer from inadequate sample size, and this suggests using daily rather than monthly data. Modeling dynamics/lags is critical in daily models, and apparently this is the first such study to model lags correctly using a general to specific approach. This is also the first risk return study to apply Wu tests for possible problems of endogeneity/measurement error for the risk variable. Results indicate a statistically significant positive relation between expected returns and risk, as is predicted by capital asset pricing models. Development of the Wu test leads naturally into a model relating aggregate risk of returns to economic variables from the risk return study. This is the first such model to include lags in variables based on a general to specific methodology and to include covariances of such variables. I also derive coefficient links between such models and risk-return models, so in theory these models are more closely related than has been realized in past literature. Empirical results for the daily model are consistent with theory and indicate that the economic and financial variables explain a substantial part of variation in daily risk of returns. The first section of this essay also investigates at a theoretical and empirical level several alternative index number approaches for aggregating multivariate risk over stocks. The empirical results indicate that these indexes are highly correlated for this data set, so only the simplest indexes are used in the remainder of the essay.

Obtaining the distribution of the profit and loss (PL) of a portfolio is a key problem in market risk measurement. However, existing methods, such as those based on the Normal distribution, and historical simulation methods, which use empirical distribution of risk factors, face difficulties in dealing with at least one of the following three problems: describing the distributional properties of risk factors appropriately (description problem); deriving distributions of risk factors with time horizon longer than one day (time aggregation problem); and deriving the distribution of the PL given the distributional properties of the risk factors (risk aggregation problem). Here, we show that expansion methods can provide reasonable solutions to all three problems. Expansion methods approximate a probability density function by a sum of orthogonal polynomials multiplied by an associated weight function. One of the most important advantages of expansion methods is that they only require moments of the target distribution up to some order to obtain an approximation. Therefore they have the potential to be applied in a wide range of situations, including in attempts to solve the three problems listed above. On the other hand, it is also known that expansions lack robustness: they often exhibit unignorable negative density and their approximation quality can be extremely poor. This limits applications of expansion methods in existing studies. In this thesis, we firstly develop techniques to provide robustness, with which expansion methods result in a practical approximation quality in a wider range of examples than investigated to date. Specifically, we investigate three techniques: standardisation, use of Laguerre expansion and optimisation. Standardisation applies expansion methods to a variable which is transformed so that its first and second moments are the same as those of the weight function. Use of Laguerre expansions applies those expansions to a risk factor so that heavy tails can be captured better. Optimisation considers expansions with coefficients of polynomials optimised so that the difference between the approximation and the target distribution is minimised with respect to mean integrated squared error. We show, by numerical examples using data sets of stock index returns and log differences of implied volatility, and GARCH models, that expansions with our techniques are more robust than conventional expansion methods. As such, marginal distributions of risk factors can be approximated by expansion methods. This solves a part of the description problem: the information on the marginal distributions of risk factors can be summarised by their moments. Then we show that the dependence structure among risk factors can be summarised in terms of their cross-moments. This solves the other part of the description problem. We also use the fact that moments of risk factors can be aggregated using their moments and cross-moments, to show that expansion methods can be applied to both the time and risk aggregation problems. Furthermore, we introduce expansion methods for multivariate distributions, which can also be used to approximate conditional expectations and copula densities by rational functions.

This diploma thesis "Methods of risk aggregation on financial markets" introduces all kinds of risk that are present on the financial markets. In the first part there are explained the ways and methods of measurement of these risks. Next there are shown the methods of aggregation of credit, market and operational risks. One of these methods are copula functions which are constructed in practical part of this thesis.

Das vorliegende Skript ist aus einer Lehrveranstaltung hervorgegangen, die von mir mehrere Jahre an der Fakultät Wirtschaftswissenschaften der TU Dresden gehalten wurde. Diese Lehrveranstaltung hatte erst die Bezeichnung "Monetäre Risikomaße" und später "Risikomaße". Mehrere frühere Fassungen dieses Skripts, das häufig überarbeitet und erweitert wurde, trugen den Namen Monetäre Risikomaße (Auflagen 1 bis 7). Die einzelnen Kapitel enthalten in der Regel die drei abschließenden Abschnitte "Übungsaufgaben", "Beweise" und "Ergänzung und Vertiefung" mit Material zum jeweiligen Kapitel, das nicht in der Vorlesung vorgetragen wurde.

Made available in DSpace on 2010-04-20T21:00:03Z (GMT). No. of bitstreams: 4 Fabio Nunes Barja.pdf.jpg: 2268 bytes, checksum: 3ff1d834aa6bf6835efe6dd07835326c (MD5) Fabio Nunes Barja.pdf.txt: 144718 bytes, checksum: 86d161817200706186fc58ff0c7b1cd4 (MD5) license.txt: 4712 bytes, checksum: 4dea6f7333914d9740702a2deb2db217 (MD5) Fabio Nunes Barja.pdf: 920446 bytes, checksum: 547cbd04aa1355d329be524b2fe94b1d (MD5) Previous issue date: 2009-11-30T00:00:00Z / Este trabalho aplica a teoria de cópulas à mensuração do risco de mercado, através do cálculo do Value at Risk (VaR). A função de cópula oferece uma maior flexibilidade para a agregação de riscos quando comparada com abordagens tradicionais de mensuração de risco. A teoria de cópulas permite a utilização de distribuições de probabilidade diferentes da normal para a modelagem individual dos fatores de risco. Além disso, diferentes estruturas de associação entre eles podem ser aplicadas sem que restrições sejam impostas às suas distribuições. Dessa forma, premissas como a normalidade conjunta dos retornos e a linearidade na dependência entre fatores de risco podem ser dispensadas, possibilitando a correta modelagem de eventos conjuntos extremos e de assimetria na relação de dependência. Após a apresentação dos principais conceitos associados ao tema, um modelo de cópula foi desenvolvido para o cálculo do VaR de três carteiras, expostas aos mercados brasileiros cambial e acionário. Em seguida, a sua precisão foi comparada com a das metodologias tradicionais delta-normal e de simulação histórica. Os resultados mostraram que o modelo baseado na teoria de cópulas foi superior aos tradicionais na previsão de eventos extremos, representados pelo VaR 99%. No caso do VaR 95%, o modelo delta-normal apresentou o melhor desempenho. Finalmente, foi possível concluir que o estudo da teoria de cópulas é de grande relevância para a gestão de riscos financeiros. Fica a sugestão de que variações do modelo de VaR desenvolvido neste trabalho sejam testadas, e que esta teoria seja também aplicada à gestão de outros riscos, como o de crédito, operacional, e até mesmo o risco integrado. / This study applies the theory of copulas to the measurement of market risk by doing the Value at Risk (VaR) calculation. The copula function offers a greater flexibility to aggregate the risks as compared to traditional approaches of risk measurement. The theory of copulas enables the use of probability distributions different from the normal to the individual modeling of risk factors. Furthermore, different association structures between them can be applied with no restrictions being imposed to its distributions. Thus, premises such as joint normality of returns and linearity in the dependence between risk factors can be dismissed, what enables the correct modelling of extreme joint events and of asymmetry in the dependence relation. After presenting the main concepts associated to the theme, a copula model was developed in order to calculate the VaR for three portfolios which are exposed to the Brazilian foreign exchange and stock markets. Afterwards, its accuracy was compared with that of traditional methodologies, i.e., delta-normal and historic simulation. The results showed that the model based on the theory of copulas was superior to the traditional ones at forecasting extreme events, which are represented by VaR 99%. When it comes to VaR 95%, the delta-normal model presented the best results. Finally, it was possible to conclude that the theory of copulas study is of great relevance to financial risks management. For further research, a suggestion is testing variations of the VaR model developed in this work, as well as applying this theory to managing other risks, such as credit, operational or even integrated risk.

Cette thèse porte sur l'étude de la modélisation et estimation de la dépendance des portefeuilles de risques et l'estimation du risque agrégé. Dans le Chapitre 2, nous proposons une nouvelle méthode pour estimer les quantiles de haut niveau pour une somme de risques. Elle est basée sur l'estimation du rapport entre la VaR de la somme et la VaR du maximum des risques. Nous utilisons des résultats sur les fonctions à variation régulière. Nous comparons l'efficacité de notre méthode avec quelques estimations basées sur la théorie des valeurs extrêmes, sur plusieurs modèles. Notre méthode donne de bons résultats lors de l'approximation de la VaR à des niveaux élevés lorsque les risques sont fortement dépendants et au moins l'un des risques est à queue épaisse. Dans le Chapitre 3, nous proposons une procédure d'estimation pour la distribution d'un risque agrégé basée sur la copule échiquier. Elle permet d'obtenir de bonnes estimations à partir d'un petit échantillon de la loi multivariée et une connaissance complète des lois marginales. Cette situation est réaliste pour de nombreuses applications. Les estimations peuvent être améliorées en incluant dans la copule échiquier des informations supplémentaires (sur la loi d'un sous-vecteur ou sur des probabilités extrêmes). Notre approche est illustrée par des exemples numériques. Finalement, dans le Chapitre 4, nous proposons un estimateur de la mesure spectrale basé sur l'estimation à noyau de la densité de la mesure spectrale d'une distribution à variation régulière bivariée. Une extension de notre méthode permet d'estimer la mesure spectrale discrète. Certaines propriétés de convergence sont obtenues / This thesis comprises three essays on estimation methods for the dependence between risks and its aggregation. In the first essay we propose a new method to estimate high level quantiles of sums of risks. It is based on the estimation of the ratio between the VaR (or TVaR) of the sum and the VaR (or TVaR) of the maximum of the risks. We use results on regularly varying functions. We compare the efficiency of our method with classical ones, on several models. Our method gives good results when approximating the VaR or TVaR in high levels on strongly dependent risks where at least one of the risks is heavy tailed. In the second essay we propose an estimation procedure for the distribution of an aggregated risk based on the checkerboard copula. It allows to get good estimations from a (quite) small sample of the multivariate law and a full knowledge of the marginal laws. This situation is realistic for many applications. Estimations may be improved by including in the checkerboard copula some additional information (on the law of a sub-vector or on extreme probabilities). Our approach is illustrated by numerical examples. In the third essay we propose a kernel based estimator for the spectral measure density of a bivariate distribution with regular variation. An extension of our method allows to estimate discrete spectral measures. Some convergence properties are obtained

Das vorliegende Skript ist aus einer Lehrveranstaltung hervorgegangen, die von mir mehrere Jahre an der Fakultät Wirtschaftswissenschaften der TU Dresden gehalten wurde. Diese Lehrveranstaltung hatte erst die Bezeichnung "Monetäre Risikomaße" und später "Risikomaße". Mehrere frühere Fassungen dieses Skripts, das häufig überarbeitet und erweitert wurde, trugen den Namen Monetäre Risikomaße (Auflagen 1 bis 7). Die einzelnen Kapitel enthalten in der Regel die drei abschließenden Abschnitte "Übungsaufgaben", "Beweise" und "Ergänzung und Vertiefung" mit Material zum jeweiligen Kapitel, das nicht in der Vorlesung vorgetragen wurde.