How AI agents could become investing’s crash test dummies

To test the effects of car collisions on the human body, researchers in the 1940s allowed themselves to be jolted to a stop on high-speed sleds, showered with broken glass and pummelled with heavy objects.

Those scientists must have cheered the invention of “Sierra Sam”, the first crash test dummy, in 1949.

Practitioners in investing – a world equally concerned with crashes, though of a different type – may soon use Sierra Sams of their own.

Some in the industry reckon artificial intelligence agents could help investment firms test foundational parts of how they operate.

“How do we structure research? What should be the output? What are the questions we should ask? How do we synthesise the information? Trying out different set-ups is hard in the real world,” says Bernd Wuebben, who oversees systematic investing and quantitative fixed income research at AB.

The agents in question are variants of so-called large language models, the same models used in chatbots such as ChatGPT.

Agentic models, though, are empowered with the ability to carry out tasks. Researchers often train them for specific jobs or prompt the models to ‘reason’ in certain ways.

Agents, the thinking goes, could allow firms to test more such variations, without incurring bruises along the way

The questions the agents might help answer are questions that define how many investment firms gain an edge over rivals, namely how firms are set up. Some in the industry call this organisational alpha.

Millennium Management is famous, for example, for running decentralised pods cut off from communicating with one another, and for imposing strict drawdown limits. Citadel is known for its intense focus on centralised risk management, making sure its portfolio managers take only idiosyncratic bets.

Which is a better system? Nobody can know for sure. And the same kinds of architectural choices face all investment organisations. Agents, the thinking goes, could allow firms to test more such variations, without incurring bruises along the way.

The idea would be to simulate investing with different configurations of agents taking on the roles of analysts, portfolio managers, risk managers, and so on, and different mixes of rules governing how the agents interact.

This is similar in spirit to experiments already underway to build investment firms entirely using agentic AI.

Wuebben has drawn up a list of possible exercises — seven, so far.

Simulations, he explains, might allow investment firms to test the benefits and drawbacks of centralised versus decentralised decision-making, or greater or lesser degrees of specialisation, or to examine quantitative compared with discretionary investment processes — all without hiring anybody new.

In the centralisation-decentralisation example, the agentic firm would simulate a set-up where portfolio managers made final decisions based on analyst recommendations, versus one in which analysts were able to act directly on ideas.

The researchers would track the speed of decision making, the frequency of bad calls, the incidence of conflicting signals from analysts and PMs, and, of course, risk-adjusted returns.

“By running these simulations, we can explore how different configurations affect overall performance, risk management and adaptability in changing market environments,” Wuebben writes in a note shared with Risk.net that sets out how the various experiments could work.

Another simulation would compare a quant-first approach — quants generate signals that humans refine — versus a human-first structure in which analysts come up with ideas for quants to test and validate. The exercise could help determine the optimal balance between quant signals and qualitative judgment, Wuebben thinks.

Yet another experiment would test different incentive structures. And yet another would run simulations under stress scenarios, to observe how organisational configurations might perform during market shocks.

Firms could learn a lot about themselves. “Understanding which set-ups perform best under stress is essential for effective risk management,” Wuebben says in the note.

Of course, for simulations to work the AI agents will need to be of similar skill to their human counterparts. Poor performance would undermine the value of the experiments. Some, though, in the industry believe such performance to be within reach.

Human judgment will doubtless always play a role in the organisational decisions that firms make. The Wayne State University professor who pioneered impact biometrics said his time as a human test dummy was vital to understanding the mathematics of crashes.

Investing Sierra Sams, though, may prove to be an unexpected use case for agentic AI.