Finance-Insurance
Time: 11.00 am
Date:05th of February 2026
Room 3001

Gero JUNIKE (LMU, Munich) “Accuracy estimation of neural networks by extreme value theory, validation of generative AI and applications to finance”

Abstract : In the first part of the talk, we consider a random variable Y_k that depends on a parameter k. For example, in a financial application, Y_k could represent the payoff of a financial contract under a certain model. The parameter k describe the contract and the model. For each k, we are interested in the expectation of Y_k, which can be interpreted as a price. This expectation can be obtained using Monte Carlo methods or Fourier techniques. To increase computational speed, researchers proposed using a neural network to learn the map f(k) := E[Y_k]. This is possible because neural networks can approximate any continuous function on a compact set. We propose using extreme value theory to quantify large values of the approximation error by the neural network. Large errors are typically relevant in financial applications. In the second part of the talk, we will not assume that the random variable of interest is given by a model anymore. We are only given M realizations. Using generative AI, we learn the distribution of the random variable to generate new samples. A possible application is determining the value-at-risk of a financial portfolio. We discuss the unwanted memorization effect (overfitting) of generative AI. We introduce a novel memorization ratio to detect this effect and apply it to an autoencoder-based scenario generator.

Organizers:  Jean-Michel ZAKOIAN & Christian FRANCQ