Contributions in Merlin

We extend the original form of prospect theory by Kahneman and Tversky from finite lotteries to arbitrary probability distributions, using an approximation method based on weak convergence. The resulting formula is computationally easier than the corresponding formula for cumulative prospect theory and makes it possible to use prospect theory in future applications in economics and finance. Moreover, we suggest a method how to incorporate a crucial step of the “editing phase” into prospect theory and to remove in this way the discontinuity of the original model.

À l’été 2007, les problèmes liés à la dette subprime aux États-Unis ont entraîné des perturbations sur de nombreux segments du système financier, en particulier sur les marchés monétaires interbancaires, obligeant les banques centrales américaine et européenne à intervenir à maintes reprises afin de rétablir un bon fonctionnement. Cet article examine les circonstances dans lesquelles une pénurie de liquidité peut apparaître et évalue différentes possibilités qui s’offrent au prêteur en dernier ressort pour restaurer la stabilité financière. Il montre également que l’évaluation des risques des entités financières à levier financier ne doit pas reposer uniquement sur les données de bilan, mais prendre aussi en compte, de manière explicite, les sûretés, l’illiquidité et l’indisponibilité potentielle des prix du marché. Nous en tirons principalement deux conclusions. Premièrement, nous établissons une hiérarchie claire entre les instruments de politique. D’après la relation entre risque et efficience, les injections de liquidité ciblées (facilités d’urgence) sont à privilégier. En effet, lorsque la liquidité est utilisée dans un but spéculatif en période de crise, les opérations d’open market non discriminatoires risquent d’attirer des participants manquant de fonds, qui peuvent détourner la monnaie centrale et en priver ceux qui en ont le plus besoin. Les injections de liquidité ciblées deviennent alors strictement préférables. Deuxièmement, à notre avis, les cessions forcées d’actifs peuvent perturber les marchés dans le cas où les investisseurs ont un levier financier élevé. Compte non tenu du financement externe et de la renégociabilité des contrats de prêt, si un investisseur au levier total est touché par un choc sur la liquidité, il sera contraint de se défaire d’une partie de ses actifs. Sur des marchés qui ne sont pas parfaitement liquides, ces liquidations induisent des baisses de prix, qui, en présence d’entraves à la gestion standard du risque, entraînent un réexamen des bilans évalués à la valeur de marché, des appels de marge et des cessions supplémentaires. Dans le pire des scénarios, l’investisseur à fort levier ne pourra peut-être pas faire face à toutes ces contractions de la liquidité et aux appels de marge dont elles s’accompagnent. Il en résulte alors un effondrement du marché des actifs illiquides, ce qui rend la valorisation de ces actifs relativement ambiguë. Pour l’investisseur, en raison de la rupture potentielle des échanges, le niveau des pertes déclenchant le défaut opérationnel est probablement plus bas que celui donné par les mesures standard du risque.

Adaptions of applications for the individual user are often not or only insufficiently developed. If an application is adapted, then usually only for certain countries and groups of people. Not often enough the effort is made to create an individualized adaption. The reasons are on the one hand the high expenditure of the realisation and on the other hand the problem to know, how such a personified adaption should look like, thus which individual needs and requirements the user has. This paper describes the implementation of a Web platform named CUMOWeb, which demonstrates an approach of an individual adaption, based on a Todo application. The site-elements are modularly built and can therefore be freely combined. Therefore an individual adaption can be automatically generated for each user. This generation is based on user specific cultural dimensions, which are provided by the ontology CUMO. In consequence it is possible to present an individually adapted interface already at the users first visit.

This diploma thesis was carried out in the context of the UBS environment. It focuses on an enmeshed software effort estimation method, which is dependent upon an integrated requirements definition, resource planning and project management. The diploma thesis discusses general estimation principles and approaches which lead to a more accurate and precise effort estimation. It points out the differentiation between business targets, effort estimation as work and duration and discusses the current situation at UBS before stressing the main problems and issues. The target situation presents a light iterative estimation approach based on experience-value-driven estimating factors. The approach separates effort from duration and focuses onto enmeshed project management activities. The effort estimation discipline consists of estimation attributes, which are mainly input and parameter definitions, the estimation model, which is the algorithmic approach and the estimation process, which defines management activities. Since the diploma thesis was written in the context of a practical implementation, the realization is part of it too, exposing hurdles to be taken. An analysis chapter discusses the set targets, potential risks and gives an outlook. This article answers the questions if and how UBS applies its effort estimation process in an adequate way. It is shown, that effort estimation may not be separated from surrounding and directly dependent processes. It presents a modular estimation approach grounded in estimating factors.

In complex software projects a lot of information about defect, release and source code history is gathered. Researchers figured out that mining these software repositories could provide valuable information about the software development. So far, software repositories were mined with traditional data mining methods which are suitable for propositional data. Propositional data is flat and homogeneous, held in a single-table-database. This thesis compares the traditional approach with relational data mining methods which are able to handle heterogeneous data. First, an introduction about relational data mining is given and then a few relational data mining tools are introduced. In a next step we present the data for our experiments and the necessary data preparations. Finally, we conduct several experiments which show the advantages as well as the weaknesses of the relational approach.

Anomaly Detection Systems have been applied successfully to a great variety of applications. To the present day there is still no system available that detects anomalies in call patterns in a Voice over IP system in order to prevent possible abuses. This work analyses di erent established methods for anomaly detection and examines their applicability in this context. The goal of this work is to design, develop and prototypically implement an Anomaly Detection System which is able to monitor a user's call behavior and detect anomalies in real time. The Anomaly Detection System developed in this work takes into account the call parameters destination number, day of the week and time of day. The pro le creation, as well as the classi cation process are realized with statistical methods. The implementation is done in C++ and connected to Asterisk® using Asterisk's FastAGI protocol. The evaluation shows that the prototype can operate in real time successfully. The false positive and the false negative rates depend on the actual values of the classi er settings (thresholds, number of calls used for pro le creation, etc.). The results show that using the same values for all the pro les do not lead to optimal classi cation results for all pro les. Further investigations with respect to a dynamic adjustment of the con guration values to the user pro le are necessary. Also, the expansion of the model to take into account additional parameters (for example the location of the user) must be considered. Therefore the model could be expanded to a multi-stage classi er.

This Student Project provides several aspects of a more efficient neighborsearch needed in particle based fluidsimulations. One of those aspects uses memorizing of its local spatial environment (environmental knowledge). This environmental knowledge can be used at a later time to calculate the neighbors. Thus neighborsearch from scratch can be avoided. Other provided aspects do not deal with the environmental knowledge to improve neighborsearch. All aspects are not competitive to each other

Measuring IP traffic traces gets increasingly more complex as data flows tend to increase over time. Processing and storage may no longer only be done by a single unit, such as a router, but may be handled by multiple units. Therefore this thesis develops two kinds of distributed storage strategies for IP traffic traces, and responds to the need of joining multiple nodes into a processing and storage network on top of one of the most used P2P frameworks, Freepastry. The first storage strategy adopts a random approach towards distributing either processing power or storage capacity. The second strategy is based on a data-centric replication strategy and is therefore completely different from the random approach. The thesis will also compare these two strategies and concludes why one strategy is favourable over the other. This work was started on the basis of an existing project from a previous semester thesis. Therefore the analysis chapter will focus on the existing project where as the following conclusion will show what parts of that application might be reusable and what parts should be replaced for a variety of reasons. Furthermore this thesis will include work from an ongoing project at the University of Konstanz where they build and test an XML based storage system called TreeTank. The implementation of this storage system will hopefully result in a more robust and highly efficient storage and processing platform for IP traffic traces. The solution to the overall problem is an architecture design called DIPStorage.

null

null