Completing Partial Implied Vol Surfaces with Variational Autoencoders
Implied volatility surfaces are ubiquitous in quantitative finance and many valuation tools use them as inputs. However, surfaces produced from market data are usually incomplete, so they need to be interpolated and extrapolated. In this talk, we will explain how variational autoencoders can remove human bias from this procedure and let the data speak for itself through unsupervised learning. The resulting light-weight models can be used to produce synthetic volatility surfaces that are indistinguishable from those observed in the market. This allows us to robustly complete partially observed volatility surfaces without making assumptions about the process driving the underlying asset or the shape of the surface.
Presenter: Maxime Bergeron: Director of Research & Development, Riskfuel
Maxime Bergeron is the Director of Research and Development at Riskfuel, a capital markets focused startup that is developing ultra-fast AI-based valuation technologies. There, his work is focused on applied machine learning and the topology of high dimensional data. Prior to joining Riskfuel, he was a faculty member at the University of Chicago. He holds a PhD in Mathematics from the University of British Columbia.
Riskfuel is pioneering the use of deep neural networks (DNNs) to accelerate the proprietary financial models that are used to calculate the values and the risk sensitivities of derivatives portfolios. Given the size of these portfolios and the many different risk sensitivities required, these models are run millions of times each day, typically in large, overnight batch processes spread over thousands of servers. Riskfuel accelerates these models, making them a million times faster so that what once took all night to run can now be completed in seconds. With Riskfuel model acceleration, you get real-time valuation and risk management … and a massive reduction in the compute workload, saving money and reducing the firm’s environmental footprint.