The event of a default for low-default portfolios, such as sovereign debt or banks, have received much attention as a result of the increasing instabilities in financial markets. The lack of sufficient default information on low-default portfolios complicates the protection of such portfolios. Default protections have typically, in the past, relied on extreme value theory and reporting the value at risk. The focus here, is the application of an engineering concept, accelerated test techniques, to the problem of insufficient data on low-default portfolios. In the application, high-default portfolios serve as stressed cases of low-default portfolios. Since high-default portfolios have more data available, viewing it as a stressed case of a low-default portfolio enables us to extrapolate the data to the low-default portfolio environment, and do estimation such as estimating the default probability for a low-default portfolio. The flexible framework through which the above is achieved, is provided. / Dissertation (MSc)--University of Pretoria, 2016. / Statistics / MSc / Unrestricted

Present thesis deals with the multivariate extreme value theory. First, concepts of modelling block maxima and threshold excesses in the univariate case are reviewed. In the multivariate setting the point process approach is chosen to model dependence. The dependence structure of multivariate extremes is provided by a spectral measure or an exponent function. Models for asymptotically dependent variables are provided. A construction principle from Ballani and Schlather (2011) is discussed. Based on this discussion the pairwise beta model introduced by Cooley et al. (2010) is modified to provide higher flexibility. Models are applied to data from nine hydrological stations from northern Moravia previously analysed by Jarušková (2009). Usage of the new pairwise beta model is justified as it brought a substantial improvement of log-likelihood. Models are also compared with Bayesian model selection introduced by Sabourin et al. (2013). Powered by TCPDF (www.tcpdf.org)

The contributions of this thesis have mainly a dual purpose: introducing several multivariate statistical methodologies where in the major of the cases only stationary of the random variables is assumed, and also highlight some of the applied problems in risk management where extreme value theory may play a role. Mostly every chapter is selfcontained, they have its own more detailed introduction and short conclusion. / Die Kontributionen von dieser Dissertation haben ein doppeltes Ziel: die Darstellung von vielen multivariaten statistischen Verfahren, wobei in der Mehrheit der Fälle nur Stationarität von den Zufallsvariablen angenommen wurde, und die Anwendungen in Risikomanagement in welchem Extremwerttheorie eine wichtige Rolle spielen könnte. Die Struktur der Arbeit ist eigenständig, mit einer detaillierten Einführung und kurzen Zusammenfassung in jedem Kapitel.

Value-at-Risk (VaR) has long been the standard risk measure in financial risk management. However, VaR suffers from critical shortcomings as a risk measure when it comes to quantifying the most severe risks, which was made especially apparent during the financial crisis of 2007–2008. An alternative risk measure addressing the shortcomings of VaR known as expected shortfall (ES) is gaining popularity and is set to replace VaR as the standard measure of financial risk. This thesis introduces how extreme value theory can be applied in estimating ES using an unconditional peaks-over-threshold method. This includes giving an introduction to the theoretical foundations of the method. An application of this method is also performed on five different assets. These assets are chosen to serve as a proxy for the more broad asset classes of equity, fixed income, currencies, commodities and cryptocurrencies. In terms of ES, we find that cryptocurrencies is the riskiest asset and fixed income the safest.

Nous présentons et étudions des méthodes d’apprentissage non-supervisé de phénomènes extrêmes multivariés en grande dimension. Dans le cas où chacune des distributions marginales d’un vecteur aléatoire est à queue lourde, l’étude de son comportement dans les régions extrêmes (i.e. loin de l’origine) ne peut plus se faire via les méthodes usuelles qui supposent une moyenne et une variance finies. La théorie des valeurs extrêmes offre alors un cadre adapté à cette étude, en donnant notamment une base théorique à la réduction de dimension à travers la mesure angulaire. La thèse s’articule autour de deux grandes étapes : - Réduire la dimension du problème en trouvant un résumé de la structure de dépendance dans les régions extrêmes. Cette étape vise en particulier à trouver les sous-groupes de composantes étant susceptible de dépasser un seuil élevé de façon simultané. - Modéliser la mesure angulaire par une densité de mélange qui suit une structure de dépendance déterminée à l’avance. Ces deux étapes permettent notamment de développer des méthodes de classification non-supervisée à travers la construction d’une matrice de similarité pour les points extrêmes. / We present and study unsupervised learning methods of multivariate extreme phenomena in high-dimension. Considering a random vector on which each marginal is heavy-tailed, the study of its behavior in extreme regions is no longer possible via usual methods that involve finite means and variances. Multivariate extreme value theory provides an adapted framework to this study. In particular it gives theoretical basis to dimension reduction through the angular measure. The thesis is divided in two main part: - Reduce the dimension by finding a simplified dependence structure in extreme regions. This step aim at recover subgroups of features that are likely to exceed large thresholds simultaneously. - Model the angular measure with a mixture distribution that follows a predefined dependence structure. These steps allow to develop new clustering methods for extreme points in high dimension.

For many real-world problems, optimization could only be formulated with partial information or subject to uncertainty due to reasons such as data measurement error, model misspecification, or that the formulation depends on the non-stationary future. It thus often requires one to make decisions without knowing the problem's full picture. This dissertation considers the robust optimization framework—a worst-case perspective—to characterize uncertainty as feasible regions and optimize over the worst possible scenarios. Two applications in this worst-case perspective are discussed: stochastic estimation and rare-event simulation. Chapters 2 and 3 discuss a min-max framework to enhance existing estimators for simulation problems that involve a bias-variance tradeoff. Biased stochastic estimators, such as finite-differences for noisy gradient estimation, often contain parameters that need to be properly chosen to balance impacts from the bias and the variance. While the optimal order of these parameters in terms of the simulation budget can be readily established, the precise best values depend on model characteristics that are typically unknown in advance. We introduce a framework to construct new classes of estimators, based on judicious combinations of simulation runs on sequences of tuning parameter values, such that the estimators consistently outperform a given tuning parameter choice in the conventional approach, regardless of the unknown model characteristics. We argue the outperformance via what we call the asymptotic minimax risk ratio, obtained by minimizing the worst-case asymptotic ratio between the mean square errors of our estimators and the conventional one, where the worst case is over any possible values of the model unknowns. In particular, when the minimax ratio is less than 1, the calibrated estimator is guaranteed to perform better asymptotically. We identify this minimax ratio for general classes of weighted estimators and the regimes where this ratio is less than 1. Moreover, we show that the best weighting scheme is characterized by a sum of two components with distinct decay rates. We explain how this arises from bias-variance balancing that combats the adversarial selection of the model constants, which can be analyzed via a tractable reformulation of a non-convex optimization problem. Chapters 4 and 5 discuss extreme event estimation using a distributionally robust optimization framework. Conventional methods for extreme event estimation rely on well-chosen parametric models asymptotically justified from extreme value theory (EVT). These methods, while powerful and theoretically grounded, could however encounter difficult bias-variance tradeoffs that exacerbates especially when data size is too small, deteriorating the reliability of the tail estimation. The chapters study a framework based on the recently surging literature of distributionally robust optimization. This approach can be viewed as a nonparametric alternative to conventional EVT, by imposing general shape belief on the tail instead of parametric assumption and using worst-case optimization as a resolution to handle the nonparametric uncertainty. We explain how this approach bypasses the bias-variance tradeoff in EVT. On the other hand, we face a conservativeness-variance tradeoff which we describe how to tackle. We also demonstrate computational tools for the involved optimization problems and compare our performance with conventional EVT across a range of numerical examples.

This dissertation proposes new statistical methods for modeling and inference for two specific types of time series: integer-valued data and multivariate nonstationary extreme data. We rely on the class of integer-valued autoregressive (INAR) processes for the former, proposing a novel, flexible and elegant way of modeling count phenomena. As for the latter, we are interested in the human brain and its multi-channel electroencephalogram (EEG) recordings, a natural source of extreme events. Thus, we develop new extreme value theory methods for analyzing such data, whether in modeling the conditional extremal dependence for brain connectivity or clustering extreme brain communities of EEG channels. Regarding integer-valued time series, INAR processes are generally defined by specifying the thinning operator and either the innovations or the marginal distributions. The major limitations of such processes include difficulties deriving the marginal properties and justifying the choice of the thinning operator. To overcome these drawbacks, this dissertation proposes a novel approach for building an INAR model that offers the flexibility to prespecify both marginal and innovation distributions. Thus, the thinning operator is no longer subjectively selected but is rather a direct consequence of the marginal and innovation distributions specified by the modeler. Novel INAR processes are introduced following this perspective; these processes include a model with geometric marginal and innovation distributions (Geo-INAR) and models with bounded innovations. We explore the Geo-INAR model, which is a natural alternative to the classical Poisson INAR model. The Geo-INAR process has interesting stochastic properties, such as MA($\infty$) representation, time reversibility, and closed forms for the $h$-th-order transition probabilities, which enables a natural framework to perform coherent forecasting. In the front of multivariate nonstationary extreme data, the focus lies on multi-channel epilepsy data. Epilepsy is a chronic neurological disorder affecting more than 50 million people globally. An epileptic seizure acts like a temporary shock to the neuronal system, disrupting normal electrical activity in the brain. Epilepsy is frequently diagnosed with EEGs. Current statistical approaches for analyzing EEGs use spectral and coherence analysis, which do not focus on extreme behavior in EEGs (such as bursts in amplitude), neglecting that neuronal oscillations exhibit non-Gaussian heavy-tailed probability distributions. To overcome this limitation, this dissertation proposes new approaches to characterize brain connectivity based on extremal features of EEG signals. Two extreme-valued methods to study alterations in the brain network are proposed. One method is Conex-Connect, a pioneering approach linking the extreme amplitudes of a reference EEG channel with the other channels in the brain network. The other method is Club Exco, which clusters multi-channel EEG data based on a spherical $k$-means procedure applied to the "pseudo-angles," derived from extreme amplitudes of EEG signals. Both methods provide new insights into how the brain network organizes itself during an extreme event, such as an epileptic seizure, in contrast to a baseline state.

The Weibull distribution, an extreme value distribution, is frequently used to model survival, reliability, wind speed, and other data. One reason for this is its flexibility; it can mimic various distributions like the exponential or normal. The two-parameter Weibull has a shape (γ) and scale (β) parameter. Parameter estimation has been an ongoing search to find efficient, unbiased, and minimal variance estimators. Through data analysis and simulation studies, the following three methods of estimation will be discussed and compared: maximum likelihood estimation (MLE), method of moments estimation (MME), and median rank regression (MRR). The analysis of wind speed data from the TW Daniels Experimental Forest are used for this study to test the performance and flexibility of the Weibull distribution.

Vehicle velocity distributions are of utmost relevance for the efficiency, safety, and sustainability of road traffic. Yet, due to technical limitations, they are often empirically analyzed using spatiotemporal averages. Here, we instead study a novel set of microscopic traffic data from Dresden comprising 346 million data points with a resolution of one vehicle from 145 detector sites with a particular focus on extreme events and distribution tails. By fitting q-exponential and Generalized Extreme Value distributions to the right flank of the empirical velocity distributions, we establish that their tails universally exhibit a power-law behavior with similar decay exponents. We also find that q-exponentials are best suitable to model the vast extent to which speed limit violations in the data occur. Furthermore, combining velocity and time headway distributions, we obtain estimates for free flow velocities that always exceed average velocities and sometimes even significantly exceed speed limits. Likewise, congestion effects are found to play a very minor, almost negligible role in traffic flow at the detector sites. These results provide insights into the current state of traffic in Dresden, hinting toward potentially necessary policy amendments regarding road design, speed limits, and speeding prosecution. They also reveal the potentials and limitations of the data set at hand and thereby lay the groundwork for further, more detailed traffic analyses.

Floods is a serious concern across Europe due to the enormous material damage and death toll. Of alltypes of floods, flash floods and large-scale river floods have become major natural hydrological hazardsin most countries. The city of Halmstad was chosen due to its placement on the southern west coast ofSweden, a region for which climate projections have indicated more precipitation and potential forflooding. In recent years a number of floods have also been observed with associated damages. Usingextreme value analysis on observed data these events can be interpreted in terms of return level valuesand their frequency of occurrence. The seasonal variation of the precipitation and discharge of thecatchment were analyzed based on 43-year precipitation and 25-year discharge observation data and therelationship to NAO index was investigated to give a preliminary overview of the hydrologicalconditions in Halmstad and its causes. The results showed that Halmstad was seasonally characterizedby high discharge in winter and lower discharge in summer with the highest rainfall. The effect of stormtracks represented by the NAO index on the precipitation and discharge in winter months was evident.This study focused on the analysis of extreme data of precipitation and discharge. The return levels forup to 50-year return period were estimated by GEV fitting. The estimated return level of discharge fora 50-year return period is 250 m³/s, and the return levels of precipitation for a 50-year flood was foundto be 68 mm/day. Two cases were selected from a compiled annual maxima discharge data set foranalyzing and comparing their weather conditions based on ERA5 data. The results showed that differentweather conditions do have an impact on the total rainfall, and there were similar patterns but largedifferences between ERA5 reanalysis data and observed SMHI data was also shown emphasizing theneed for long-term observational data sets and further evaluation of reanalysis data.