Projects
AIDA: Artificial Intelligence for Due-Diligence Analysis
This project develops AI-powered information retrieval and natural language processing (NLP) techniques for e-discovery and due diligence analysis on financial and legal documents. The goal is to assist legal professionals in finding critical information within large sets of disclosed documents.
- PHD Student: Madhukar Dwivedi
- Supervisors: Marc Francke (ABS), Jaap Kamps (ILLC)
- External Partners: Imprima, Zuva AI
ESG Regulation Impact on Financial Stability
- PhD Student: Matteo Valle
- Supervisors: Simon Trimborn (ASE), Debraj Roy (IvI)
- External Partner: ING
HyperMining: Explainable Anti-Money Laundering
In this project we aim to develop an innovative Anti-Money Laundering methodology using advanced AI methods using hypergraph representations and process mining which can give an integral view of the transactions involved, deal with the inherent complexity of the data, and still be understandable for the experts analyzing the data so they can substantiate their decisions.
- PhD Student: Fatemeh Gholamzadeh Nasrabadi
- Supervisors: Marcel Worring (IvI), Michael Werner (ABS)
- External Partner: Transaction Monitoring Netherlands
Knowledge-Driven Learning for XAI in Fraud Detection
In this project we aim to develop a novel neuro-symbolic framework that mainly combines the strengths of both the data-driven approaches (which comes with adaptability, autonomy, and good qualitative performance) and the knowledge-driven approaches (which comes with interpretability, maintainability, and well-understood computational characteristics) to provide explanations for experts in terms of relevant features and the structures in-between.
- PhD Student: Adia Lumadjeng
- Supervisors: Erman Acar (IvI/ILLC), Ilker Birbil (ABS)
- External Partners: Mollie, ING
Robust Fraud Detection through Causality-Inspired ML
Payment platforms like Adyen use technology to efficiently detect fraud. Fraud detection is challenging, since both the genuine and fraudulent customer behavior changes over time and across markets. Machine learning is crucial for this task, but current methods are susceptible to learning spurious correlations. The goal of this project is to leverage causality-inspired machine learning methods to improve the robustness of fraud detection methods to distribution shifts.
- PhD Student: Jakub Reha
- Supervisors: Sara Magliacane (IvI), Ana Mickovic (ABS)
- External Partner: Adyen
Systems for AI Data Quality in Finance
The impact of data errors on the output of AI models is difficult to anticipate and measure, and these errors can negatively impact regulatory compliance. Therefore, this project aims to enable non-technical users to validate and increase the quality of their data. For that, these users should be able to express data quality rules in natural language. We will design a data driven approach to leverage such rules to assist a domain expert to finetune data quality rules and “stress test” downstream AI models. This project favors a strong data engineering background combined with an interest to engage with.
- PhD Student: Yichun Wang
- Supervisors: Sebastian Schelter (IvI), Kristina Irion (IViR)
- External Partner: ABN-AMRO
Improving Machine Learning Methods for Contingent Claim Pricing and Hedging
To be able to use new machine learning techniques for applications in mathematical finance such as the pricing of contingent claims, option pricing algorithms must be designed which allow calibration, pricing and hedging to be based on characteristics of financial time series which summarize important features of large datasets without specifying a detailed model. The proposed PhD project aims to contribute to this aim, by focusing on new mathematical methods to characterize signals (such as, for example, path signatures) and particularly challenging applications for such methods in mathematical finance (such as forward prices in energy markets and regime switching).
- PhD Student: Wouter Andringa
- Supervisors: Asma Khedher (KdVI), Drona Kandhai (KDVI/IvI), Michel Vellekoop (FEB)
- External partner: EY
Data-Driven Mathematical Modeling and Computing for Pricing and Hedging Renewable Energy Contracts
The increasing share of renewable energy sources is driving a transformation of energy markets toward a more sustainable and green paradigm. However, this transition introduces the need to address new types of risks, including volumetric, grid, and climate risks. This project focuses on the pricing and hedging of renewable energy contracts influenced by these risk factors, with a particular emphasis on computational methodologies. By leveraging advanced mathematical modeling and data-driven approaches, the research aims to develop robust strategies for managing these emerging risks effectively.
- PhD Student: Konstantinos Chatziandreou
- Supervisors: Sven Karbach (KDVI/IVI), Drona Kandhai (KDVI/IvI)
- External partner: -
Reinforcement learning for integrated transaction authentication and authorization
Transaction authentication and authorization requires different decisions to be made that are currently supported by various models acting in a pipeline. A more integrated approach could make better informed decisions, but requires complex high-dimensional optimization. We aim to support this optimization using reinforcement learning techniques, where we expect to tackle challenges such as a high-dimensional action space, off-line learning and evaluation, and trading off multiple objectives.
- Student:Adi Watzman
- Supervisor: Herke van Hoof (IvI)
- External Partner: Adyen
Learning causal drivers for user behavior in transaction pipelines through reusing existing data or by designing new A/B tests
Many of the current steps in transaction approval pipelines aim at modelling the causal drivers of the user behavior (e.g. what authentication/authorization option would cause a reduction in friction for the payments for specific sets of users) and estimating the causal effects of certain choices. One way of estimating these causal effects is through A/B tests on a specific set of options, which can be expensive and only cover a small sample of users with certain features. After these A/B tests are performed they are usually not reused or combined with previous tests or observational results. In this project we aim to reuse existing data from A different sets of A/B tests, as well as design more focuses, accurate and cheaper experiments to learn causal drivers of user behavior reliably and robustly.
- Supervisor: Sara Magliacane (IvI)
- Student:Roel Huisman
- External Partner: Adyen
Mechanistic Interpretability for Time-Series Transformers
The success of transformers in processing sequential data for language and vision tasks (e.g., ChatGPT, DALL-E) has sparked significant interest in applying these models to forecasting. Consequently, numerous research proposals have emerged, drawing considerable attention and prompting both discussion and skepticism regarding the true capabilities of these inherently black-box models. This project seeks to address these concerns by adopting a mechanistic interpretability approach—an approach that has proven valuable in understanding language and vision models. By doing so, we aim to offer deeper insights into the internal workings of time-series transformers.
- PhD Student: Angela van Sprang
- Supervisors: Erman Acar (IvI/ILLC), Jelle Zuidema (ILLC)