Call for Submissions 2025

Paul Hager

Paul Hager

Project: Fast Robust Control for Data Streams with Memory

We increasingly face high-frequency time series data: streams of observations arriving rapidly, often irregularly, and strongly shaped by past behavior. Finance is a prime example. Modern markets are driven by high-frequency and algorithmic trading and by fragmented, decentralized venues, creating new challenges for pricing and hedging financial risk and for calibrating models to data. A central difficulty is that market dynamics have memory: what happens next depends not only on the current price, but also on the recent history of trading activity, volatility, and order flow. Many machine-learning approaches tackle this with large "black-box" models trained end-to-end. While powerful, such models can be hard to interpret, fragile on irregular data (as in limit-order books), and offer limited analytical insight.

A structured alternative is based on ideas from modern mathematics known as rough path theory. Its central practical tool is the signature of a time series: a compact set of features that summarizes the information contained in the full history of the series. Signatures are stable when data are observed at different time resolutions and are expressive enough to represent a wide range of path-dependent effects.

This project initiates research at the intersection of reinforcement learning and signature-based, non-Markovian control: how to learn robust, fast decision rules when actions must react to rapidly arriving market data with strong memory effects. Reinforcement learning is increasingly important for sequential decision-making under uncertainty, but most standard methods rely on simplified "state" descriptions that break down when the past matters. By using signatures to encode relevant history in a stable and compact way, the project aims to make learning-based control both more reliable and more transparent in high-frequency settings.

Leopold Zehetner

Leopold Zehetner

Project: Data-Driven Strain Engineering for Improved and Cost-Efcient CAR–T Virus Manufacturing

CAR-T cell therapy has transformed cancer treatment by using a patient's own immune cells to fight malignant tumors. However, the therapy remains prohibitively expensive—costing between €50,000 and €450,000 per patient. A major cost driver is the production of retroviral vectors, the specialized "delivery vehicles" that insert therapeutic genes into T cells. Current production costs can reach €100,000 per batch, limiting access to these life-saving treatments worldwide. This seed project addresses this challenge by improving retroviral vector production through genetic optimization of the murine cell lines used in manufacturing. Over decades of cultivation, these cell lines accumulate genetic changes that can reduce production efficiency. By understanding and correcting these genetic variations, we can significantly boost viral yields and reduce manufacturing costs. The project focuses on two key cell lines essential for retroviral vector production: NIH-3T3 and PG13. We will generate high-quality genome sequences for both cell lines using cutting-edge sequencing technology, map genetic variations and identify areas of instability that limit production efficiency, update computational models of cellular metabolism to pinpoint metabolic bottlenecks, and develop preliminary hypotheses about how to optimize these cells for better viral production. These foundational results will provide essential proof-of-concept for a larger follow-up research proposal aimed at implementing targeted genetic engineering strategies. Even independently, the genomic data and computational models will become valuable resources for the biotechnology community, advancing CAR-T manufacturing globally. Ultimately, this work aims to reduce costs and expand patient access to transformative cancer therapies.