Florian Ofenheimer-Krach

Abstract

Time series forecasting is an omnipresent topic in many application areas. Real world time series forecasting problems are often characterised by a complex data structure with discrete, irregular, noisy observations with missing values, posing a notoriously difficult problem for classical approaches as well as machine learning models. In this talk, we introduce the Neural Jump ODE framework, which is an artificial neural network-based machine learning model, designed to deal with the complex setting of time series forecasting problems. Importantly, we prove for this fully data-driven model that it converges to the optimal prediction given by the conditional expectation. Different versions of this model are proposed for problems with noisy observations, for long-term predictions and for input-output systems, which can all be combined if necessary. 

Our speaker

From September 2019 until February 2025, I did my PhD studies with Prof. Josef Teichmann in D-MATH at ETH Zurich. Starting from March 2025, I am a PostDoc in his research group.

To become a member of the Rough Path Interest Group, register here for free.