Research Themes

HIPERFIT comprises a number of research themes relevant to computing in finance, and execution on high-performance backends. The following ones have been identified to guide and connect together our activities in the center.

Risk Scenarios

We try to describe the transition from observables (like current prices and historical data) to scenario generation and from scenario generation to reporting and management. In order to, for example, pick a quantile scenario on the basis of high-dimensional data, one has to address modeling aspects, correlation and other dependency aspects, and extreme value aspects. Risk scenarios play a main role in solvency issues for generation of different types of capital requirements. These requirements are fed back into management processes and lead to decision making. Within this theme we address the mutuality in the risk scenario generation and decision making processes. The topic calls upon advances in data structures and numerical methods for so-called differential games.

Capital requirements based on risk scenarios are applied in valuation, for instance in the so-called cost of capital method. Within this theme we are interested in the relation between arbitrage free valuation and cost of capital valuation generated in a given solvency regime.

Model Specification

How much detail is necessary and how much is sufficient in a financial model? Going from one extreme, namely model-free relative valuation (some prices are given completely in terms of other prices) towards another extreme, namely advanced multi-dimensional jump-diffusions processes with a high-dimensional parameter set, what are the benefits and costs for the given application in mind, be it solvency issues, accounting issues, or management issues? A main part of this is, in the first place, to study these advanced models in details and clarify their strengths and weaknesses, also in relation to their optimisation-precision trade-off on high-performance systems.

DSLs for Financial Information

Declarative domain-specific languages (DSLs), operating at a high level of (programming) abstraction, are already widely-used in the financial sector to describe a range of financial products (derivatives). We want to design similarly expressive and automatically optimizable declarative specification languages for other financial information, and especially for financial models (see above).

Our goal is a DSL framework with broad application coverage, that enables fast implementation, easy maintenance, extensibility and reusability across financial institutions, in short: low total cost of ownership (TCO). Financial information applications should cover computing applications such as reporting to auditors and public authorities, data communication with clearing houses, internal reporting and statistics, computations for the purpose of internal risk management, and flexible integration for standard routines such as accounting and confirmation processing.

Extracting Parallelism from Declarative Specifications

The core theme of this theme is the rendition of large-scale financial computations into a form that exposes inherent parallelism. To this end, we aim to develop a language that can be productively used to express large-scale numeric computations in a natural style, close to familiar mathematical notation, and yet is, at the same time, efficiently executable on a variety of modern parallel platforms.

The paradigm of data-parallelism (by G.Blelloch) appears to be a good match for the HIPERFIT application domain: it enables concise and long-term maintainable specifications of a wide variety of inherently parallelizable computations, without committing to any particular implementation strategy or execution environment. We will design a suitable such language, with a compositional formal semantics and cost model, to enable correctness proofs and performance estimates.

The DSL development for financial models will lead to additional or modified requirements at a later stage of the project. Typical operations for valuation of stochastic financial models need to be translated into the provided parallel operations. Necessary and useful abstractions and patterns of parallelism will be identified from working on concrete projects.

High-Performance Backends for Novel Hardware

These are activities to ensure that the models, in their executable form, can execute on next-generation processors. The research will focus on mapping the parallelism that is expressed by the functional programming activities onto a number of parallel computer architectures. The systems activities will work in close collaboration with other projects that seek to exploit next-generation processors.

In a first stage, we will investigate a number of current algorithms from computational finance IT for execution on next-generation processors. Based on these experiences, a preliminary backend for next-generation-processors will be integrated with the existing solutions from partners; before integrating the full backend with the engine for functional programming from the project itself.