Scalable GPU Accelerated Computing for Risk-Neutral Modeling
Master thesis about High Performance Computing in finance using GPU’s
Meet Alex
I first started at Ortec Finance as a student-assistant looking to gain some working experience in my field of studies, Computational Science. After about 5 months, I was looking for an interesting research subject for my master thesis. As it turned out, Ortec Finance itself had a great line up of research projects from which I chose a project considering high performance computing with GPUs. This research subject was by far one of the most interesting ones I had seen so far. Furthermore, I also wanted to continue my part-time job here at Ortec Finance, for which they provided me the option to combine both working and graduating. I kept working two days a week as a student-assistant at the Research department, while spending the remaining three on my thesis project. This dual construct I can recommend to any student who is allowed to extend the schedule for his/her project a little.
About the project
The thesis considered the design of a scalable high performance computing framework using Graphical Processing Units (GPUs) for the acceleration of Monte-Carlo and nested Monte-Carlo simulations (simulations within simulations). Monte-Carlo simulations are a computationally intensive but flexible and straightforward method to simulate future market trajectories and value even very complex derivative contracts by generating scenarios. Because each scenario path is independent from all others, these simulations can be greatly accelerated by generating each path in parallel (at the same time). This allows the computation of risk measures (e.g. Value at Risk) and calibration of risk-neutral models within reasonable amounts of time without the need for more advanced mathematical methods. The goal was to extend existing implementations to be able to handle multiple correlated underlying assets and freely scale the number of GPUs and parallel CPU processes that offloaded work to those GPUs depending on the workload of simulation. In the end we achieved great results in improving the scalability and performance of the framework. Plans for even further extensions can already be found on the Ortec Finance website in the form of a new available research project, where we are looking to tap into the limitless resources of the cloud.
For me, working on this thesis project was a great experience. I received great supervision at Ortec Finance with invaluable weekly feedback sessions and many great suggestions by colleagues from many different departments. Unsurprisingly from my glaring positive review, after my graduation I started working at Ortec Finance fulltime as a Quantitative Developer within the EFIS Department.
Interested?
Alex’ thesis is not publically available, one can request a copy via techlabs@ortec-finance.com.
Used tools & Hardware: CUDA, Python, Scipy, PyCuda, GTX Titan X GPU’s
Author: Alex