Learning jump-diffusion dynamics from irregularly sampled data
Tim Jahn (TU Berlin)
| When |
Jun 30, 2026
from 01:00 to 02:00 |
|---|---|
| Where | M7 101 |
| Attendees |
Tim Jahn |
| Add event to calendar |
vCal
iCal
|
Tim Jahn
Title: Learning jump-diffusion dynamics from irregularly sampled data
Abstract: Modeling continuous-time systems from irregular observations is central to applications ranging from clinical monitoring to high-frequency finance. While neural SDEs offer a flexible tool, they struggle with path dependence and the computational costs of solver-based training. We introduce a trajectory generator matching framework for jump-diffusion processes with memory, extending flow-matching beyond purely continuous dynamics. Our approach combines a neural drift with stochastic delay terms for path dependence and a learned rate kernel for state- and time-dependent jump intensities. By deriving closed-form generators for conditional jump-diffusion bridges, we obtain a simulation-free loss function that bypasses unstable backpropagation through solvers.
Website: https://www.tu.berlin/en/math/data-science/jahn
In Room M7 101, 1st floor, Monod campus, ENSL