Financial expert demands: AI must not only be clever, but also secure

Artificial intelligence (AI) is rapidly transforming the financial world – because its maturity is rapidly evolving and its potential applications are extremely diverse. Here are three examples of current, common applications in asset management: AI can, among other things, read thousands of business and financial texts every day – stock market reports, analyst commentaries, company announcements, and more – and assess how new information will affect individual stocks.

In addition, AI can observe how prices on the capital markets develop in real time, thus recognizing how investors are currently evaluating news. A third example: AI can improve the quality of traditional portfolio models by deriving return expectations not only from the past but also from current trends and forecasts.

Artificial intelligence can already do a lot today. Yet the financial world is only at the beginning of a gigantic AI boom: The number of applications will continue to grow rapidly for years to come. While word of the benefits of AI in asset management has spread more slowly than in other industries, they are now well known to investment firms.

Here are three examples: AI processes vast amounts of data in seconds – much faster than teams of analysts ever could. AI recognizes patterns and relationships that are too complex or even completely invisible for humans. And it considers significantly more influencing factors than traditional investment approaches – and can therefore make more informed investment suggestions.

Opportunities and risks are closely linked

However, asset managers naturally know from their day-to-day investment practice that where there are opportunities, there are also risks. And this also applies to the use of AI. One of the risks in this area is manipulation – for example, when language models (LLMs) interpret seemingly harmless text differently, or in the worst case, incorrectly, through rewording it in company reports. And then algorithms or humans make decisions based on such incorrect AI-generated interpretations.

Other risks arise from the interplay of technologies that learn from each other, adapt to each other, and automatically develop behavioral patterns. These behavioral patterns can distort the market and thus be harmful to competition. It is not impossible that events like the flash crash on May 6, 2010, could repeat themselves: At that time, the S&P 500 stock index fell by almost six percent within a few minutes; the Dow Jones Industrial Average Index even lost more than nine percent. Stock trading volume rapidly increased sixfold. And numerous stocks temporarily fell to a fraction of their previous price.

People often don't even notice other technological effects, even when the consequences are significantly negative – for example, for liquidity (i.e., the tradability of securities, bid-ask spreads, and price variances during order execution), price stability, and trust in market functionality. Trust is the foundation for any allocation of client funds.

Making the financial sector more transparent and secure

The level of trust in all types of technological processes depends heavily on how transparently they function and the extent to which users understand their workings. More and more scientific institutions are devoting themselves to both aspects—functional transparency and functional understanding—in the field of artificial intelligence.

Because these aspects are so important for (financial) technical and ethical reasons, the Liechtenstein asset manager Plexus Investments has been offering the "Award for Artificial Intelligence in Finance," endowed with €10,000, since 2020. The jury intensively reviewed the research papers submitted in the fifth year of the competition (2024) – primarily master's theses and dissertations – in the first half of 2025. The jury consisted of:

• Walter Farkas , Professor of Quantitative Finance at the Institute of Banking and Finance at the University of Zurich and associate member of the Department of Mathematics at the Swiss Federal Institute of Technology (ETH) Zurich,
• Christof Kutscher , Co-founder and Executive Chair of Lancry Natural Capital (Zurich), and
• Stefan Mittnik , Professor of Financial Econometrics, until 2020 holder of the chair of the same name at the Ludwig Maximilian University of Munich and co-founder of the digital financial services provider Scalable Capital.

The jury recently selected two winners. Both examined subtle but significant risks inherent in AI models. And both share the same vision: to make AI for the financial industry not only smarter but also more secure.

Language model weaknesses clearly revealed

The winner – Aysun Can Türetken from the University of Zurich – holds a master's degree in economics (minor in data science). The title of her master's thesis, submitted for the award, is "An Adversarial Attack Approach on Financial LLMs Driven by Embedding-Similarity Optimization."

She says: "The increasing use of natural language processing (NLP) models in the financial world raises questions: for example, whether NLP models are reliable or robust enough for the challenges of asset management. My research goal was to better understand the models and identify weaknesses so that we can develop more robust and reliable models."

For her master's thesis, Aysun Can Türetken investigated how even the smallest reformulations in financial texts can mislead language models to the point where the models even change their sentiment classifications for the texts. She had GPT-4o generate sentences that were almost identical or only slightly different from each other – with the requirement to maintain their meaning and implied sentiment.

The sentences were supposed to remain the same in terms of content. The language models tested were FinBERT and FinGPT, which are smaller and cheaper than GPT-4o and specialize in financial texts. Several investment firms now use FinBERT and FinGPT for automatic sentiment classification. These models were supposed to classify the marginally changed sentences. And what happened?

For example, FinBERT classified the sentence "The company reported a net loss of 50 million euros in the third quarter, down from a profit of 20 million in the same period last year" as "negative." Correct! In contrast, FinBERT classified the only minimally modified sentence "The company had a net loss of 50 million euros in Q3, compared to 20 million euros profit a year earlier" as "neutral." Incorrect! The research of this year's AI award winner Aysun Can Türetkens clearly demonstrated in numerous tests that AI-based sentiment models are susceptible to manipulation.

Unintentional unfairness proven

The second award winner, Wei Xiong, a doctoral student at the University of Oxford, focused on a different aspect of risk: the behavior of learning AI agents in financial markets. AI agents are systems that independently pursue goals specified by humans, such as fair price determination. Wei Xiong's dissertation is titled "Dynamics of Market Making Algorithms in Dealer Markets." For his dissertation, he modeled how algorithmic market makers (autonomous price setters) interact with each other when they are – in effect – in competition with each other.

His simulations showed that even without direct communication, such agents can develop pricing strategies that appear to be tacitly coordinated with each other, resulting in permanently elevated spreads. "Algorithms don't need to talk to each other to behave as agreed," says Wei Xiong. "This form of emergent coordination is a new systemic risk."

Developers should therefore think beyond pure performance optimization and document and monitor technological learning processes. "This allows them to determine, among other things, whether algorithmic market makers are inadvertently 'collusing,'" adds Wei Xiong. Regulators are also facing new challenges, as existing rules such as best execution and anti-collusion rules are designed for explicit human collusion – not for self-organized patterns through algorithmic learning. "Greater transparency, audits of trained models, and a more considered design of price request processes can reduce such risks early on," says Wei Xiong.

A small anniversary of great achievements in AI research

The research work of Wei Xiong and Aysun Can Türetken demonstrates once again that excellence in AI in asset management not only means relatively high forecasting accuracy, but also transparency, security, and trust. 'Once more,' in the context of the AI ​​​​promotion award, means: in its fifth year.

Optimize forecasts, manage liabilities, calculate option returns

In light of this small anniversary, it's worth taking a look back: Last year, three award winners impressed the jury with their research submitted for 2023. Valentin Hasner (now Junior Portfolio Manager Fixed Income Frontier Markets at Azimut Investments in Luxembourg) investigated the predictive power of neural networks for European equity markets based on long-short-term memory (LSTM) networks—with differentiated results across different time periods and market phases.

Konrad Müller (now Associate Quantitative Researcher at JP Morgan Chase in London) demonstrated with "Deep ALM" how asset-liability management (ALM) can be optimized through deep reinforcement learning – in a regulatory-compliant and strategically robust manner. Mathis Mörke (now Assistant Professor at the École supérieure de commerce de Paris Business School in Paris) impressed with a series of papers on options market research, in which he used complex machine learning models, among other things, to quantify the predictability and mispricing of individual options – opening up new perspectives for factor and return models.

Improving bond yield forecasts and statistical learning

Two award winners were honored in 2023: Sebastian Ott (now Associate Private Markets at Liechtenstein asset management company Principal) and Urban Ulrych (now Postdoctoral Researcher at ETH Zurich) demonstrated with their research in 2022 that machine learning not only improves forecasts but also has a profound impact on the practice of portfolio management and derivatives valuation. Sebastian Ott analyzed the cross-section of US corporate bonds and demonstrated that machine learning identifies yield factors that not only increase portfolio returns but also reduce drawdown risk. Urban Ulrych combined statistical learning with challenging financial practice: from dynamic currency hedging and stable asset allocation to accelerated pricing of American options using neural networks.

Systematic bond strategies and evasive corporate leaders

The 2024 award winners

In 2022, two AI scientists brought two previously neglected facets into focus in their research papers submitted in 2021: Colin Glag (now Associate Portfolio Manager Fixed Income at Allianz Global Investors in Frankfurt) demonstrated how machine learning—particularly decision tree-based models—can be used to predict corporate bond returns and develop systematic investment strategies.

And Sasan Mansouri (now Assistant Professor for Digitalization and AI at the University of Groningen) developed a model that identifies so-called “non-answers” ​​in analyst calls – and demonstrated their economic relevance: Evasive answers from company management lead to negative price reactions and increased uncertainty in the market.

Identify equity risks and improve factor modeling

And last but not least: In 2020, Benjamin Moritz and Riccardo Tegazi demonstrated the diverse ways in which AI-based methods can contribute to asset pricing: Benjamin Moritz (now Head of Investment Research at the Frankfurt-based family office Finvia) combined text-based economic indicators with machine learning to better capture risks and returns in the stock market.

Riccardo Tegazi (now a trader specializing in European government bonds at Bank of America in Paris) demonstrated in a cross-country comparison how machine learning models – from XGBoost to neural networks – can significantly increase the predictive power and explainability of factor modeling.

Looking back, it is clear that artificial intelligence is making a significant contribution to the advancement of asset management, is developing into a highly relevant research area, and the Plexus Investments "Award for Artificial Intelligence in Finance" has become a benchmark for excellent and responsible research at the interface of AI and finance.

About the guest authors:

Günter Jäger founded Plexus Investments in 2006 and has served as Managing Director of the Liechtenstein-based investment company ever since. He initiated Plexus Investments' AI think tank in 2017.

Aysun Can Türetken holds a Master’s degree from the University of Zurich and is the author of the research paper “An Adversarial Attack Approach on Financial LLMs Driven by Embedding-Similarity Optimisation”, published in 2024 and awarded the AI ​​Prize in 2025.

Wei Xiong holds a PhD from the University of Oxford and is the author of the dissertation “Dynamics of Market Making Algorithms in Dealer Markets,” published in 2024 and awarded the AI ​​Prize in 2025.