Why the person behind the AI model remains crucial

AI and machine learning have long been part of everyday practice in asset management. However, the more readily available algorithms become, the more important it is to consider who understands them, uses them correctly and scrutinises them critically. Carsten Rother, Co-Head of Research Forecasts, and Dr Desislava Rakova, Research Forecasts – who recently undertook further training as part of a machine learning programme at Stanford University – discuss why further training, an understanding of models and human judgement remain crucial – and how Quoniam deploys AI where it can create genuine added value.

Why is continuing professional development in AI and machine learning so important for research teams today?

Desislava: A sound academic education forms the foundation. But in a field that is developing as rapidly as artificial intelligence, it is not enough to acquire knowledge once and then assume that it will suffice in the long term. Many methods, tools and applications are constantly changing – and some have already found their way into practice before they are fully integrated into university curricula.

At the same time, it has become very easy to apply algorithms these days. There are software packages, cloud infrastructure and AI assistants that quickly generate suggestions. You can feed data into a model and get a result very quickly. But that is precisely where the risk lies: just because a model is running doesn’t mean you understand whether it is suitable for your specific research question.

My further training therefore focused primarily on gaining a deeper understanding of what goes on behind the scenes in these models: How are algorithms parameterised? Why does a model sometimes fail to converge? Which settings are appropriate for which problem? And when does a result merely look good, but is in fact not robust?

Carsten, why is this deeper understanding so important for Quoniam?

Carsten: Because quantitative asset management is very much about the finer details. I like to compare it to Formula 1. It’s not enough just to have a good car. You have to understand and optimise every single component: the data, the models, the parameterisation, the validation, the implementation.

Running algorithms is becoming easier and easier. The added value comes from knowing which algorithm to use and when, what its characteristics are, and why it works – or doesn’t work – in a particular case.

Then machine learning is no longer a black box. To someone who only sees the output, it may seem that way. But if you understand the mechanics, the parameters and the limitations, it ceases to be guesswork and becomes a conscious research process.

Where does this manifest itself specifically in the investment process?

Desislava: Algorithms can play a role at various stages. They can be incorporated into forecasting models, generate signals or help assess risks more accurately. One example from the fixed-income sector is downgrade models. These aren’t about directly forecasting returns, but rather about determining whether a company’s credit profile might deteriorate.

A vivid example from the course was the so-called ‘cocktail party problem’. You have a recording with many voices and background noise, and an algorithm is supposed to isolate the voice of the actual speaker. At first glance, this has nothing to do with asset management, but the analogy is very clear: we, too, try to filter out the relevant signal from the noise of the markets.

What do you need to understand before selecting a model?

Desislava: It always starts with defining the problem. What am I actually trying to solve? Is it about forecasting returns, assessing risk, classification, or gaining a better understanding of a market mechanism? That determines which model makes sense in the first place.

If you skip this step, you can very quickly waste a lot of time. A model may run technically but still fail to deliver meaningful output. In that case, however, the problem may not lie with the algorithm, but rather with the fact that the question was unclear or the parameterisation was incorrect.

You talk a lot about parameterisation and fine-tuning. Why are these adjustments so crucial?

Carsten: Because that’s often where the real difference lies. The big idea is important, but it’s not enough. What matters is how the model is configured, which data is used, which training and validation periods are chosen, and how the results are interpreted.

This doesn’t just apply to machine learning. Even with traditional factors such as value or momentum, the details make a huge difference. How do I define value? How do I measure momentum? How do I deal with outliers? How do I test whether an effect is stable? This detailed work is an essential part of proprietary research.

What does robustness mean in this context?

Carsten: Robust means that a model must not just look good on historical data. It must also be able to function in future situations it has not yet encountered. This is particularly important in asset management, because no two crises are the same.

Take downgrades, for example. A downgrade can have many causes: debt, rising interest rates, a changed business model or other factors. There isn’t one perfect threshold that applies to every company. That’s why we aren’t looking for a model that explains the past as well as possible but is too specialised. We’re looking for robust solutions that remain resilient even in different market environments.

Has the course changed your day-to-day work?

Desislava: Yes. I’ve become more confident when I need to structure a problem and find a solution. At the same time, I’ve become more critical – even towards my own projects or those of my colleagues.

Nowadays, I ask more frequently: Why was this particular model chosen? Why these parameters? Why this training period? Why this validation? These are simple questions, but they are crucial. Data preparation is important, but model selection and parameterisation are just as important.

What role does AI play in day-to-day research?

Desislava: I definitely see AI as a tool. For repetitive or very well-structured tasks, it’s already delivering significant efficiency gains. It can help us work faster, prepare code, structure ideas or identify initial solutions.

But when it comes to more challenging research questions, a plausible-sounding answer is not enough. You need to know very precisely what you’re actually expecting, which data is relevant and how a result should be interpreted. If this groundwork is missing, AI can even lead to a lot of time being wasted – or to relying on a solution that only appears convincing on the surface.

In asset management in particular, the aim is not simply to obtain averagely good answers. We need to develop robust, transparent and investable solutions. Human judgement remains crucial for this.

Carsten, what does this mean for the research team?

Carsten: This culture of critical thinking is central to what we do. We want people who question academic research, our own models and even AI-generated suggestions. That is precisely why we support training programmes such as this course. They benefit not just one person, but the whole team.

Quantitative asset management combines capital market expertise, mathematical understanding and technical implementation. AI can make us more efficient. But when it comes to the crucial details, we still need experienced researchers.

This also aligns with our fundamental understanding of AI: we want to use it where it makes sense and creates real added value. But it must be embedded within a controlled, transparent investment process.

What is the key message?

Desislava: A sound academic background is the foundation. But in a field that is developing so rapidly, you have to keep learning. You need to understand what’s going on behind the algorithms and mustn’t rely on a tool to automatically deliver the right solution.

Carsten: Exactly. Applying algorithms is becoming easier and easier. The key is to understand why they work – and when they don’t.

Conclusion

AI can make research faster and more efficient. However, the decisive added value arises when people understand models, recognise their limitations and critically interpret the results. For Quoniam, human expertise therefore remains a central component of a robust and transparent investment process.