Robeco fires up AI thematic ETF

Dutch asset manager Robeco has launched an exchange-traded fund (ETF) that uses artificial intelligence to pick investing themes in stocks.

The launch is part of the fund manager’s broader push into active ETFs, a field in which global assets increased since 2017 from $58 billion to $565 billion last year.

Robeco’s $80 billion quantitative investing unit teamed up with experts from its thematic investment group to develop, test and calibrate the firm’s Dynamic Theme ETF, launched in October.

The firm is using natural language processing, the form of AI models that interpret language, to detect and switch in and out of emerging investment themes, as well as determine which stocks are theme beneficiaries and losers.

“The goal is to invest in themes a bit earlier than the rest of the market,” says Mike Chen, Robeco’s head of next generation research. Thematic investing, by definition, is high conviction long-term investing – where human emotion might cloud investors’ judgement. This makes it the perfect candidate for AI, he says.

We use a variety of natural language processing to identify the thematic universe

Mike Chen

Robeco uses topic detection algorithms to tease out potential investment themes from companies’ own pronouncements, including conference call transcripts, company descriptions, and so on.

Once themes are identified, the firm employs a supervised learning clustering model to sort the investable universe of stocks into groups that align with theme labels that Robeco defines.

Finally, large language models are used again, but this time to conduct sentiment analysis to determine which companies are winners and losers within each theme. The algo detects, for example, that Nvidia tends to be talked about positively in the context of the growing market for GPU silicon chips, and Intel negatively.

“We use a variety of natural language processing to identify the thematic universe, cluster [stocks], and identify who benefits and who gets hurt by the themes,” Chen says.

One theme that AI detected is the digitisation of restaurants, which began during Covid when the industry was forced to move rapidly online. Some companies found they could improve efficiency if they digitised, allowing them to better control their inventory, develop new menu items and anticipate when a certain product might be in demand.

As well as the NLP models, Robeco applies a more conventional quant factor filter to ensure the ETF favours companies that are reasonable picks within each theme. The ETF would avoid theme winners if they were highly overvalued, for example.

It’s been a very tough year for thematic investors, so we’ve done relatively well

Mike Chen

Robeco began building its ETF two years ago. The concept is relatively simple, Chen says, but there were many nuances to figure out, such as how to construct clean theme clusters, how to assign themes to companies, and whether one company should be included in multiple themes.

While the AI has free rein in detecting themes and the corresponding winners and losers, Robeco has put in place human oversight, too, to ensure the machines don’t make mistakes.

“The machine is statistical,” Chen says. “Hopefully it gets right more than [it gets] wrong. But it could detect a theme that was nonsense. So we have a fundamental portfolio manager on this product as well.”

Overall, he says, the exercise has shown how hard it is for humans to think of new investable themes. By contrast, a portfolio manager shown a list of 100 themes generated by the AI can easily determine which ones make sense, he says.

Robeco launched the strategy internally as a fund towards the end of 2023, to build up a one-year track record, before launching the ETF for outside investors in October. At the time of writing the ETF was up 86 basis points since its launch. Robeco’s internal fund version of the strategy was up 1% since inception and 2% year to date versus its benchmark, the MSCI World index.

“It’s been a very tough year for thematic investors, so we’ve done relatively well,” Chen says.