AQR Experiments With GPT to Improve Returns

In back tests, trading with any of these large language models, including ChatGPT, would have helped investors outperform the market from 2004 to 2019, according to AQR’s Bryan Kelly.

Illustration by II

Illustration by II

The ascent of ChatGPT has prompted quantitative asset manager AQR to further explore the role of generative artificial intelligence in improving investment returns.

The chatbot is a poor predictor of future returns because — having been trained only once in 2021 — it has no ability to process real-time stock market data. (It’s also been programmed to not give investment advice.)

At the same time, markets incorporate information quickly, if not instantaneously.

“In finance, the problem is very different,” said Bryan Kelly, head of machine learning at AQR and a finance professor at Yale School of Management. “By and large, markets are efficient because they are extremely competitive. The efficiency eliminates the predictability in returns. Any information that’s out there and available, it’s impounded into prices very quickly and makes future price movements very difficult to predict.”

Still, some parts of ChatGPT’s language model can be applied to investments and used to improve the portfolio construction process, he said.

Kelly explained that it’s possible to harness some predictive information from text in publicly available financial information, such as analyst notes and news, to build portfolios.


Here’s how: Generative AI tools work in roughly two steps. The first is the compression stage, where it transforms general text into what’s called ‘lower-dimensional representation,’ which summarizes the meaning of the text. The second is the prediction stage, where AI translates the compressed ‘meaning’ into a prediction for subsequent text. The compression stage is particularly useful for predicting investment returns, according to Kelly.

“What we do is that we [take] the compression step, [which] goes from some document, like a news article about a particular stock, and then [transform] it into a more meaningful, smaller numerical representation of what that article is talking about,” Kelly said. “That can be useful for us.”

Put simply, generative AI tools like ChatGPT have allowed portfolio managers to process news or other financial documents more efficiently. “We capture the meaning from a whole bunch of different financial-related text documents, and then we intercept these document-level representations,” Kelly said. “We’ll pull them outside the GPT and plug those representations into our models, [which have] things that we care about on the finance and portfolio management side.”

Kelly has been experimenting with other generative AI tools as well, including Meta’s OPT and Google’s BERT, since 2017.

In back tests, trading with any of these large language models, including ChatGPT, could have helped investors significantly outperform the market from 2004 to 2019, according to Kelly.