Eduard van Gelderen
As mentioned in previous columns, I believe the surge in artificial intelligence is unstoppable. Whereas we’ve seen many applications—for example, research assistance and the discovery of trends—most were aimed at productivity improvements: obviously of value, but not leading to disruptive innovation or system solutions.
But lately, we see fundraising by organizations that move AI to the next level: from static large language models via agentic AI to autonomous decision systems. One example is EquiLibre Technologies, a Prague-based firm that uses reinforcement-learning models to pick stocks. These models differ from the use of LLMs because they are not about information retrieval or word generation, but about sequential decisionmaking through interaction and feedback—a very different AI technique altogether. Another example is Grace Investment Machine, which basically automated the entire process of generating market intelligence, forming investment hypotheses, testing assumptions and putting ideas to work, all with a feedback loop to improve the system.
A third example, AlphaCircle AI, part of ETH FinSure Tech Hub in Zurich, provides clear thinking behind these latest developments: The AI challenge is not to come up with better predictions, but to have an adaptive system in place making allocation decisions under uncertainty. The company’s manifesto states: “The traditional model of discretionary portfolio management struggles to scale to the complexity of modern financial markets.”
The complexity is the result of expansion of data, the interconnectivity of information, and the speed and nonstationary nature of market dynamics. The time-consuming and sequential process of traditional asset allocation processes simply does not match this market environment. AlphaCircle AI’s track record is stellar, albeit short and nonverified.
I follow these new institutions with great interest, because slowly, but steadily, we are moving up the investment value chain and, by combining the different steps in the value chain, AI-based system solutions become a reality. After all, agentic AI will combine and coordinate the different steps in the value chain. The beauty is that the system will learn from its mistakes and only get better. But will it truly deliver what we expect?
Mark Steed, CIO at Arizona’s Public Safety Personnel Retirement System, puts it as follows: “We risk building our own Deep Thought—the supercomputer in the movie “Hitchhiker’s Guide to the Galaxy” that spent 7.5 million years computing the answer to everything, coming up with the number 42. It left everyone stumped, because nobody knew what the exact question was to begin with. That sums up AI worship in a nutshell. Many groups will be disappointed by AI’s outputs if they haven’t done the unglamorous work … to bring clarity to their decisions and processes.”
The SpaceX Question
The SpaceX IPO is an interesting case in this respect. For sure, CIOs must have scratched their heads looking into this IPO, asking if they should follow the index composition (where the company was added outside of standard timelines) and considering questions about the company’s valuation and governance. I wondered how the new AI initiatives would deal with this.
I soon realized that it was perhaps not a very intelligent question and that I was likely to fall in the trap of believing that AI has the right answer for every problem. After all, deep reinforcement learning needs a lot of data for the models to become effective, and the SpaceX IPO was a single event with clear idiosyncratic characteristics. There are two possible outcomes here: The systems ignored the IPO, because there was no learning yet, or they allocated to the IPO. In the case of the latter, based on what?
AlphaCircle AI’s Sebastian Owen provides the following valuable insight: “The real strength of deep reinforcement learning isn’t predicting one extraordinary event like the SpaceX IPO. It’s continuously improving capital allocation across thousands of decisions under uncertainty.”
Hence, rather than replacing human judgment, these systems shift its role—from making individual investment decisions to designing objectives, governance frameworks and institutional constraints within which autonomous allocation systems can learn and operate.
Future Role of Machine-Native Systems
This brings me back to the question of whether so-called machine-native allocation systems truly deliver. I certainly believe in AlphaCircle AI’s claim that the traditional asset allocation processes have a hard time making sense of the complexity in today’s markets. However, the practicality really depends on what type of investor you are and the readiness of your organization.
A hedge fund trying to make use of anomalies in capital markets will most likely be interested in the latest developments. Moreover, most are very comfortable with technology and digitization and, as such, are ready to move into using machine-native allocation systems. Yet anecdotal evidence suggests that the investment horizon is short-term and daily trading volumes rather high.
And that is understandable: AI models need data—a lot of data. High-frequency data is boosting the success of the AI models, but it comes with a short-term investment horizon. Therefore, if you are a long-term asset allocator whose main investment purpose is to allocate savings to the best investment opportunities in the real economy, the current developments might not help you much. The current models are hardly equipped to work with low-frequency data. Perhaps more importantly, many long-term asset allocators are not ready to reap the full benefits of AI, simply because the right data, processes and governance are not in place. They are also unwilling to make the commitment to reveal their implicit investment knowledge as input in the development of AI tools.
Grappling With Complexity
So should we, as long-term asset allocators, ignore the AI trend? I don’t think so. AI developments are moving very quickly, and (deep) reinforcement learning will become more relevant for long-term investors too.
Let me give two examples. Lately, the total portfolio approach is a hot topic: assessing each individual investment’s contribution to the total fund, agnostic of asset classes. In order to do this correctly, a lot of cross-sectional data is needed. And let’s be honest: A lot of investors struggle with the quality of their historical data at large. More specifically, getting access to the appropriate data regarding their private market investments is equally challenging. The second example is so-called three-dimensional-investing based on: risk, return and resilience. What exactly resilience means is up for debate, but it goes beyond statistical terms such as downside risk.
In both of the above examples, we need to get a grip on complex systems, defined as a network of many interacting, interdependent parts that collectively produce behaviors that cannot be predicted by simply looking at the individual components. However we ultimately define resilience, it is embedded in the complex system on which we currently hardly have any grip. And this is exactly where AI and (deep) reinforcement learning is going to pay off. The new models are the first step. Developments will catch up very quickly.
Eduard van Gelderen served as the head of research at FCLTGlobal in 2025 after spending more than six years as the CIO of PSP Investments in Montreal. Prior to his role at PSP, he worked for the investment office of the University of California and was CEO of APG Asset Management in the Netherlands. He recently launched Brave Foresight, an investment management consultancy company focusing on innovation and artificial intelligence.
This feature is to provide general information only, does not constitute legal or tax advice, and cannot be used or substituted for legal or tax advice. Any opinions of the author do not necessarily reflect the stance of CIO, ISS Stoxx or its affiliates.
Tags: Arizona Public Safety Personnel Retirement System, Artificial Intelligence, disruption, initial public offering, investment decisionmaking, Mark Steed, productivity, SpaceX