A Quant Firm Gets a Performance Boost From AI — At Least in Tests

“The machine learning-based methods are a lot better than traditional methods…They can better model risks,” says Robeco’s Patrick Houweling.


Illustration by II

Asset managers are racing to see how machine learning and artificial intelligence can improve their performance. Quant manager Robeco, which has $170 billion in assets under management, is the latest to seriously experiment with the tools and get some promising results.

In an interview with Institutional Investor, Patrick Houweling, co-head of quantitative fixed income at Robeco, said his team has been working to integrate machine learning tools into the firm’s investment process for quantitative credit. Robeco started with $5.5 billion in credit strategies that provide exposure to the value factor. Similar to value stocks, Robeco defines value in fixed income as bonds that are cheap relative to their fundamentals.

Houweling said Robeco’s results show that machine learning can significantly improve risk-adjusted returns for strategies that use the value factor. When applied to an investment-grade bond portfolio that uses a value style, for instance, the machine learning-based approach improved the information ratio — which measures the excess return per unit of risk — from 1.42 to 1.83, according to a recent paper by Houweling’s team. When applied to a value-oriented, high-yield bond portfolio, the model that used machine learning increased the information ratio from 0.99 to 1.75.

“The machine learning-based methods are a lot better than traditional methods…They can better model risks,” Houweling said.


Robeco’s exploration of machine learning comes at a time when numerous asset managers are investigating the possibility of using AI in investments and distribution. For example, Rayliant Global Advisors plans to hire more AI talent to expand its investment research capabilities. Hedge fund AQR has been experimenting with ChatGPT-like large language models to improve returns. PanAgora has also been exploring the capacity of AI tools for winning mandates.

Houweling said his team of 16 portfolio managers and researchers spent six months studying the ML-based approach. “We wanted to understand the intuition behind the method,” Houweling said. “Once we went through all these different steps of the research, we decided to make a change to accept machine learning as a new tool in our toolkit.”

But to fully execute the ML-based approach in clients’ portfolios requires an infrastructure update, which may take another one or two months, according to Houweling. “The research stage has been finished,” Houweling said. “We know exactly what we want to do and how we want to do it, but we need to implement it in the software on a daily basis to calculate value signals.”

Next, Houweling’s team plans to test the same machine learning-based approach on other factors, including momentum and size. “We want to use the same tool and apply it to other factors…because we now think we understand this tool and we like it,” he said.