ING Group

ING Group – Joint Professorship (UT–ING Collaboration)

Since May 2021, Joerg Osterrieder has held a joint professorship under the UT–ING collaboration, working alongside ING’s Global Analytics team to drive AI research in finance. The appointment covers federated learning, synthetic data, reinforcement learning, and credit-risk early-warning systems, and sits at the intersection of academic research and applied industry practice. The five-year partnership combines these two vantage points to develop AI-driven solutions in finance, covering model-risk frameworks, data analytics, and scalable business integration (ai-in-finance.eu).

Research and Education Activities

Student projects and theses form a core component of the collaboration. More than eight master’s theses have been completed at ING covering topics such as Explainable AI, credit risk, ESG integration in credit models, and sustainability-report extraction, and a current PhD project is exploring confidence scoring in large-language-model outputs within retrieval-augmented generation pipelines (ai-in-finance.eu). At the doctoral level, the collaboration guides candidates in reinforcement learning for digital finance through the MSCA DIGITAL network and the COST FinAI action, with joint training weeks and specialist workshops that bring academic and industry perspectives into direct contact. Workshops and lectures extend the reach of this work; a recent Data Analytics and Quantitative Models workshop in Amsterdam in February 2025 brought together model validation, risk management, and AI techniques including retrieval-augmented generation, risk analytics, and governance testing (ai-in-finance.eu).

Synthetic Data and LLM Integration

One technical strand of the collaboration concerns synthetic financial data and large-language-model integration. GAN- and VAE-based methods for synthetic financial-data generation have been explored as ways to enrich training datasets and to preserve privacy where access to real client data is restricted. On the language-model side, retrieval-augmented generation pipelines have been developed for document extraction with explicit attention to confidence scoring and hallucination detection, so that language-model output can feed into downstream decision processes with calibrated uncertainty rather than unqualified claims (ai-in-finance.eu). These strands aim at a practical question: how to use generative and language-based AI responsibly in a regulated bank.

Credit-Risk Forecasting and Early Warning

Credit-risk forecasting and early warning form a second technical strand, where hybrid systems combine time-series methods with classical machine-learning algorithms to flag deteriorating credit conditions before losses crystallise. Master theses supervised under the programme have examined credit-deterioration detection and retail default prediction, producing empirical evidence that the combined approach outperforms single-model baselines on the bank’s portfolios (ai-in-finance.eu). The work feeds directly into ING’s credit-risk and fraud-detection frameworks, where several of the models have now been adopted in production.

Explainable AI and Regulatory Alignment

Explainable AI forms a third strand, responding to the regulatory and operational demand for interpretability in fraud detection and default-probability models. Methods have been developed and validated for compliance and governance contexts, so that a model’s predictions can be explained in terms that risk managers and supervisors can scrutinise. The work connects to a broader research agenda on validating machine-learning models against bank-grade operational and regulatory standards.

Reinforcement Learning in Finance

A fourth strand applies reinforcement learning to dynamic portfolio-management problems. Reinforcement-learning modules taught during the MSCA training week covered Q-learning and policy-gradient methods tailored to sequential decision-making in finance, with an emphasis on realistic trading environments and the difficulties of evaluating RL agents on finite historical data (ai-in-finance.eu). The academic programme doubles as an industry training pipeline: PhD researchers and ING practitioners engage with the same methods and the same evaluation challenges, which shortens the loop between research innovation and production use.

Impact and Outcomes

The collaboration has supported more than eight master theses, with several of the resulting models now deployed in ING’s credit-risk and fraud-detection frameworks, and has contributed to PhD training programmes and professional events that strengthen the AI talent pipeline (ai-in-finance.eu). It has enabled real-world AI deployments at ING, including language-model-based data extraction and federated risk models that align with regulatory and privacy requirements. The ING–UT joint professorship integrates current AI research — federated learning, synthetic data, reinforcement learning, explainable AI, and LLM pipelines — into the bank’s risk-management and analytics infrastructure. The partnership has produced research outputs, student projects, and production AI models that strengthen ING’s AI capabilities and support a working academic–industry research ecosystem.