Keywords: high performance computing, parallel computing, financial management, derivatives pricing, path dependent instruments, portfolio optimization, stochastic programming, asynchronous algorithms, benders decomposition, parallel programming model
Astract: High Performance Computing is useful in the field of finance for solving problems which are defined on models of financial variables in the form of sequences of scenarios along with their realization probabilities. Both the evolution of stock prices and interest rates is frequently described in this manner. This work deals with the two problem classes of determining prices of financial instruments, and of determining optimal portfolios of assets, with respect to some objective function and constraints. Dynamic optimization techniques allow for multiple planning periods, whereas stochastic dynamic optimization problems take into account also probabilities and exhibit (exponentially growing) tree structures, which can become very large. Computation times for solving these problems can extend to hours and days, hence high performance computing techniques of achieving speed up are desirable.
The major approach for performance improvement in this work is parallel computing. It includes the parallel implementation of Monte Carlo simulation techniques as well as of backward induction methods for pricing path dependent interest rate derivatives, in particular constant maturity floaters with embedded options. In the optimization part, the nested Benders decomposition method of multistage stochastic optimization has been parallelized in a synchronous as well as in an asynchronous version. The parallel implementations obtain speedups ranging from reasonable to excellent and demonstrate the potential of high performance computing for financial applications. In addition, they served as case studies in the development of software tools for high performance computing within the framework of the Special Research Program No. F011 AURORA "Advanced Models, Applications and Software Systems for High Performance Computing" of the Austrian Science Fund (FWF).
The data parallel programming language HPF+, with extensions for clusters of SMPs, has been successfully employed in the implementation of pricing algorithms. A path notation has been specified as an extension to Fortran 95, allowing for the high level formulation of parallel algorithms operating on lattice structures. The parallel programming model of a distributed active tree has been designed and implemented on top of Java's threads and RMI. Parallel implementations of the nested Benders decomposition algorithm in Java demonstrate that this is a suitable language for high performance computing. The OpusJava component framework, as well as the JavaSymphony class library, and the distributed active tree model proved their usefulness as programming support environments in the implementation of parallel tree structured algorithms.
In addition to the parallelization of sequential existing algorithms, the improvement of known parallelization approaches, and the use of specialized parallel programming languages and programming models, an increase in performance has been achieved by algorithmic developments.
The generalization of the classical backward induction method allows for the faster calculation, i.e., in linear instead of exponential time, of prices of a class of instruments exhibiting "limited" path dependence, demonstrating that highly effective approaches of performance improvement combine the levels of algorithms and parallel implementation.

Keywords:

Keywords: information acquisition, Web, information extraction, workflow, dataflow, meta-search, LumberJaczk
Astract: Having the right amount and combination of the right information, available in the right time is an unquestionable advantage.
The Web, being the largest information database, has become an unavoidable source for acquiring the knowledge of these three "rights".
However, it is not available for free. It requires non-trivial data search, capture, transformation and combination of data from many unrelated services, their semantic organization, and personalization.
Such processes need be automatized and repeatable for maintaining the knowledge up-to-date. Unfortunately, the extent of relevant data sources is immense and heterogeneous and the nature of the structuring of the Web information is not suitable for automatization. The information is encoded in visual structures (e.g. HTML pages), intended for humans.
The problem of information extraction from static semi-structured sources has been the focus of the recent research. In our work, we rather aimed the complementary, behavioral aspects of information extraction (Web navigation and interaction) and the aspects of information acquisition processing models.
The main contribution is the specification of a light-weight, dynamic and agile interactive Web information acquisition processing framework.
We particularly focused on two main components of the system: a task coordination workflow and a model for data outputting. The proposed models have been successfully implemented and applied in applications.
We present the most innovative one: LumberJaczk - a light-weight system for on-top-of-Web applications, i.e. client-side, portable and very personalizable applications reusing the Web.
The proposed approach reveals several new challenging fields, such as:
decentralized information acquisition, collaborative semantic databases.

Keywords:

Keywords:

Keywords:

Keywords:

Keywords:

Keywords:

Keywords: