London Mathematical Finance PhD Day
Friday, October 17th, 2025

Distributional Adversarial Attacks and Training in Deep Hedging

We propose an adversarial training framework to enhance the robustness of deep hedging strategies against distributional shifts. By reformulating the adversarial optimization problem over a Wasserstein ball, our method efficiently learns hedging strategies resilient to market perturbations and demonstrate superior out-of-sample performance.

A Gibbs Sampler for Financial Network Models with multiple CCPs

We consider a network reconstruction problem for markets with multiple central counterparties (CCPs) that are subject to bilateral netting. We propose a sampling method to reconstruct bilateral cleared amounts from aggregate information. We prove theoretical properties of the sampler, and apply it to compute the distribution of risk measures in stylized networks.

Nonparametric Estimation of Self- and Cross-Impact

Price impact refers to the empirical observation that executing a large order affects the price of a risky asset in an adverse and persistent manner. We develop an offline nonparametric estimator for concave multi-asset impact models, and implement it on both proprietary metaorder and public order flow data. This is joint work with Natascha Hey and Eyal Neuman.

Lévy-Driven Arcade Processes for Stochastic Interpolation

Arcade processes provide a method for strong stochastic interpolation between zeros at fixed times. In this talk, I present an extension whereby the driving process is allowed to exhibit jumps, leading to a construction of arcades driven by Lévy processes.

Mean-field control for heterogeneous systems of forward-backward stochastic differential equations

We study heterogeneous systems of forward–backward stochastic differential equations indexed by a type space, where the system exhibits mean-field interactions through the joint law of the states. We establish a stochastic maximum principle and a verification theorem for the associated optimal control problem.

I don't know much about utility functions, but I know what I like: Learning Optimal Collective Pension Investment and Consumption for Differing Preferences

We train a recurrent neural network to learn optimal pension investment and consumption strategies for a range of preferences. To address high variance in our utility function across parameters, we propose a novel architecture that ensures the network learns the optimum. This allows us to infer utility functions from individuals by observing the outcomes they prefer.

Risk-Sensitive Portfolio Management with Proportional Transaction Costs

We study a continuous-time risk-sensitive portfolio optimization problem under proportional transaction costs. The investor aims to maximize the long-run risk-sensitive growth rate for a non-zero risk-sensitive parameter θ. Depending on the value of θ, exactly one of five distinct controls is optimal.

Signature-Based Generative Models for Temporal Point Processes

A rough-path lift for càdlàg counting paths enables a signature-based Wasserstein GAN (SigWGAN) that extends signature methods beyond continuous paths. The resulting model achieves stable training and competitive generative quality.

Please register for the PhD day using the link below. Registration is required.

If you have any questions, please contact one of the organisers.