Sorry, you need to enable JavaScript to visit this website.
Skip directly to content

Join Peter Caspers at the New York Society Meeting and enjoy in-person networking with your fellow quant finance professionals.

Event Abstract:

In this presentation we will discuss design principles for modern quant finance libraries and illustrate them giving code examples from the popular “Open-Source Risk Engine” project.

We demonstrate how we separate the business logic from the core computational layer, how we ensure readability and maintainability of the source code and how we seamlessly integrate modern techniques like adjoint differentiation and the abundant supply of GPUs for high performance calculations.

At the heart of this lies a scripting layer and an explicit computation graph builder with novel memory management algorithms that simultaneously minimizes tape sizes for AAD allowing the processing of arbitrary sized portfolios and for efficient dynamic generation of fused kernels on GPUs. We discuss how to incorporate non - path wise operations like the omnipresent conditional expectation operator without sacrificing simplicity and performance using an implementation of “Fries’ trick”, which was a major contribution to the theory of adjoint derivatives calculation in recent years. We also discuss the limits of adjoint differentiation and alternatives for cases where it does not apply.