Dynamically Controlled Kernel Estimation

Understanding the potential of machine learning for the financial services industry is a key driver of innovation at the moment. We are pleased to announce that we are contributing to this effort with a research paper authored by our team members Jörg Kienitz and Nikolai Nowaczyk working jointly with Gordon Lee and Nancy(Qingxin) Geng.

We are revisiting the classical problem of calculating conditional expectations accurately and efficiently. This is a key challenge for many financial applications such as American option valuation, XVA or the calibration of local stochastic volatility models. We provide a novel data driven and model free approach relying on Gaussian Process Regression, kernel density estimation and control variates to improve upon the known shortcomings of the traditional LSM method. We illustrate the technique with examples from multidimensional exotic option pricing in the Heston and rough Bergomi model.

Read the full text here.