Key challenges include the fiduciary element of asset management, outdated information in some widely available models, and a preference for traditional machine learning approaches.
On the positive side, generative AI offers significant productivity gains, helping to process large amounts of data, make faster decisions, or summarize bottom-up analyzes based on reports and historical data.
Kev Toohey, principal at Atchison Consultants, noted that while large language models (LLMs) have recently gained attention, his firm has been using proprietary machine learning models as part of its investment process for more than five years.
“These models help us assess market conditions and develop signals for our tactical asset allocation perspectives.”
Additionally, Atchison Consultants uses separate machine learning models to evaluate the characteristics of individual assets and managers. These models aim to predict the investment return profile in relation to economic and market conditions.
“We view data management and visualization as a key tool for our team to effectively identify and test investment opportunities and have hired a data scientist who devotes a significant amount of resources to internal coding of our models investment and reporting,” Toohey said.
While acknowledging the role of AI in optimizing multi-asset portfolios, Toohey emphasized that his firm still relies on “deterministic models, not AI” for portfolio optimization.
“We balance the benefit of repeatability of calculations for things like portfolio optimization with remaining preferred,” he said.
“For less well-defined issues, such as where markets are in the cycle or the future performance of specific assets, we use machine learning models to inform our analysts.
Similarly, Sebastian Mullins, head of multi-asset and fixed income business at Schroders Australia, confirmed that generative AI offers process acceleration benefits.
He emphasized, however, that traditional machine learning remains an effective tool for modeling “more complex, multidimensional relationships in data.”
Mullins warned that generative AI is less useful for real-time decision-making because publicly available language models often rely on outdated information, such as ChatGPT, whose data only extends through 2021 .
“This means that generative AI is providing incorrect or outdated data or news headlines,” he said.
In contrast, he noted, machine learning is useful for tasks such as predicting future market prices, making forecasts of economic variables, as well as determining structural breaks or outliers in data.
“We currently use deep machine learning modes (using neural networks) to determine short-term interest rate forecasts relative to market prices, clustering analysis to determine interest rates. Conditional interest and currency spot probability distributions, as well as reinforcement learning to determine the fair value of government bonds. » he said.
Mike Chen, head of Next Gen Research at Robeco, noted that while AI will transform many aspects of how multi-asset portfolios are managed, some areas, such as face-to-face customer service, will continue to be. rely on human skills.
Chen also pointed out that data availability could be a limiting factor for certain asset classes.
“We believe AI has an important role to play in all assets for which there is sufficient data available. In some asset classes where data availability is not great, the role of AI will be relatively less, as AI algorithms require data to be effective,” he said.
Similarly, Toohey pointed out that asset classes with frequent market valuations, like listed stocks, are generally better suited to machine learning models than illiquid assets, which rely more on subjective or opinion-based valuations .
“We don’t give money to AI to manage it, but it allows [us] to take into account all the different data points, where before we might have taken 10, now we can take 30. And that allows you a broader reach or allows you to explore different geographies with the same level of detail but without necessarily having experts in all areas,” he said.
Boyle said he is uncomfortable relying on AI for asset classes in general at this time, but said he sees the potential for AI to improve transparency on less liquid and less transparent markets by processing data more efficiently than humans.
“I think over the next 18 months it may turn into running money, but from our perspective it’s certainly a data tool intended to improve productivity rather than actually impact the “money in different markets,” he said.
