Antoon Pelsser & Erik Winands: The risk manager's toolbox

October 9, 2025

_{This article was originally written in Dutch. This is an English translation.}

In a world full of uncertainties and risks, quantitative risk management combines traditional models and innovative AI techniques. The key to success? The right technique at the right time.

By Prof. Dr. Antoon Pelsser, Balance Sheet Management Specialist at Rabobank and Professor of Mathematical Finance & Actuarial Science at Maastricht University, and Prof. Dr. Ir. Erik Winands, Professor of Quantitative Risk Management at the University of Amsterdam and Head of Capital Adequacy & Scenario Analyses at Rabobank

The world is spinning faster than ever. Geopolitical tensions, cyber threats and climate change, among other things, are creating a more complex and unpredictable playing field. In this world full of uncertainties, risk management has become a strategic pillar for financial institutions. To keep track of all developments, risk managers are constantly expanding their quantitative toolbox. In addition to tried and tested models, new technologies are also finding their place, with artificial intelligence (AI) being the latest addition. Quantitative risk management is the field in which these techniques come together with the aim of identifying, quantifying and managing all risks: from credit and market risks to liquidity and operational risks. ^[1] Although AI offers unprecedented new possibilities, the familiar techniques have by no means lost their value. In this article, we illustrate this with two case studies. The first case study deals with a model for relocation options within a mortgage portfolio, the second with studies on portfolio selection.

Case study 1 Dutch mortgages: the lesson of the relocation option

Embedded options in Dutch residential mortgages, such as the relocation option (also known as the “transfer option”), pose an interesting challenge for the risk management of a mortgage portfolio. This option allows homeowners to transfer their existing mortgage to a new home, which is particularly advantageous when market interest rates have risen. Historically, however, little use was made of this option: between 2002 and 2022, interest rates fell almost continuously, meaning that the transfer option was rarely advantageous.

A purely data-driven model, based on machine learning for example, would come to a misleading conclusion in such a period: “No one ever transfers their mortgage.” After all, if interest rates are only falling, it is more advantageous to take out a new mortgage at the lower interest rate, and the transfer option is not interesting. But in mid-2022, the situation changed radically. Within a year, interest rates rose significantly, making the old (lower) interest rates on existing mortgages attractive when moving house. Figure 1 clearly shows that within the customer group of movers (applicants who are buying a home and whose current home is also owner-occupied), the use of transferable mortgages rose from virtually 0% to around 30% in a short period of time, and that our model is able to predict this accurately.

This highlights the danger of models that rely solely on historical data without understanding the underlying rational choices of customers. Fortunately, our modelling was not purely empirical, but based on a behavioural economic framework that considers households to be rational decision-makers at the time of moving house. This enabled us to correctly predict the sharp increase in portability behaviour, even without precedent in the data.

The lesson is clear: for embedded options, a fundamental approach based on incentives and choice behaviour is essential. Especially in a world where financial conditions can change rapidly, only such a model offers robust insights. Data-driven methods have their value, but without a theoretical framework, they can fail when it really matters.

Case 2 Portfolio selection: the lesson of the simple benchmark

The second case we discuss falls within the context of investment strategies, for which we delve into the academic literature. An influential scientific article by DeMiguel et al. ^[2] from 2009 showed that a simple 1/N strategy – the equal distribution of assets across all available investment options – regularly performs no worse than more complex strategies, unless large amounts of historical data are available. The reason for this observation is that the naive 1/N benchmark is robust to errors in the estimation of the required parameters. This conclusion is supported by studies on more recent datasets^[3]. Although these studies certainly do not recommend implementing the 1/N benchmark as an investment strategy, they do warn that blindly relying on advanced models and strategies, without considering their limitations, can be risky.

Data-driven methods have their value, but without a theoretical framework, they can fail when it really matters.

In recent months, two articles have been published in Finance Research Letters, investigating whether AI techniques such as deep reinforcement learning^[4] and large language models^[5] might be capable of structurally beating the 1/N benchmark. Although both techniques perform very well in controlled environments, they can reach their limits in practical applications. This is because they require large amounts of data to effectively learn to recognise patterns. As a result, these AI techniques are unable to consistently outperform the simple 1/N strategy for the datasets studied.

In general, three things are crucial for AI models: data, data and data. This second case also illustrates that in situations where historical data is limited and/or the past is not a reliable predictor of the future, it is not self-evident that AI and other advanced techniques always offer the most suitable solution.

Conclusion

Without a doubt, the potential of AI in quantitative risk management is very promising. AI offers new opportunities to identify dependencies, trends and early warning signals in complex datasets, enabling risks to be managed more efficiently and effectively. The temptation to use AI for every issue is therefore great. However, our two case studies have shown that this may not always be the best choice. The well-known psychologist Abraham Maslow already stated in the 1960s: if all you have is a hammer, everything looks like a nail. ^[6] Fortunately, however, we have a well-stocked toolbox at our disposal; after all, not every problem is a nail. Sometimes a traditional approach works better, sometimes AI, and sometimes a combination of the two. The power of effective risk management therefore lies not in the technology itself, but in the ability to use the right technology at the right time.

^{[1] Winands, E.M.M. (2024). Quantitative risk management: A world full of opportunities. Inaugural lecture, University of Amsterdam.
[2] DeMiguel, V., Garlappi, L., & Uppal, R. (2009). Optimal versus naive diversification: how inefficient Is the 1/N portfolio strategy? The Review of Financial Studies, 22(5), 1915–1953.
[3] Gelmini, M., & Uberti, P. (2024). The equally weighted portfolio still remains a challenging benchmark. International Economics, 179, 100525.
[4] Kruthof, G., & Müller, S. (2025). Can deep reinforcement learning beat 1/N. Finance Research Letters, 75, 106866.
[5] Perlin, M.S., Foguesatto, C.R., Müller, F.M., & Righi, M.B. (2025). Can AI beat a naive portfolio? An experiment with anonymised data. Finance Research Letters, 78, 107126.
[6] Wikipedia (2025). https://en.wikipedia.org/wiki/Abraham_Maslow. Retrieved on 01-07–2025.}