Bridgewater Associates, one of the world’s largest hedge funds, has been an early adopter of artificial intelligence and machine learning. In particular, the firm is now experimenting with large language models like GPT-3 to enhance its investment process.

Co-CIO Greg Jensen has long been passionate about the potential of AI, even investing early in prominent research lab OpenAI. So it’s no surprise that Bridgewater is on the cutting edge of applying large language models to finance.

In a recent interview on the Odd Lots podcast, Bridgewater co-CIO Greg Jensen explained how the firm is utilizing these advanced AI systems. According to Jensen, large language models have some key strengths that make them useful for investing:

  • They have read and “understand” massive amounts of text data. This gives them a broad knowledge base that can be tapped to generate investment ideas and theories.
  • They can rapidly produce huge numbers of potential investment theses at scale. Whereas human analysts are limited in how many ideas they can conceive and test, large language models can instantly generate thousands of hypotheticals.
  • They can interpret and summarize complex statistical model outputs. This allows them to “dialogue” with quantitative systems and effectively communicate insights.

However, Jensen notes large language models also have significant weaknesses:

  • They are prone to “hallucinations” - making up answers that seem plausible but are totally inaccurate. This is because they are focused on producing coherent text, not factual accuracy.
  • They have no real understanding of time or the ability to forecast regime changes. Without special tuning, they simply regurgitate relationships from historical data.

To mitigate these problems, Bridgewater uses a hybrid approach:

  1. Large language models hypothesize potential market theses and investment ideas.
  2. Quantitative models test these theories against historical data, assessing their past predictive power.
  3. Human analysts evaluate the results, probing for weaknesses in the logic or areas where markets may behave differently than in the past.
  4. Promising ideas are refined and retested, iteratively strengthening the investment thesis over multiple cycles.

To further strengthen its AI capabilities, Bridgewater has partnered with Elemental Cognition to develop a “reasoning engine” that works alongside large language models. This system is designed to mimic human logic and reasoning abilities, in order to refine the raw text produced by language models.

Jensen explains the reasoning engine is intended to overlay logical rules and structure on top of the unstructured text from language models. This allows it to identify flawed logic, unlikely conclusions, or areas where the language model may be fabricating information.

The reasoning engine has capabilities to ask questions, probe the reasoning behind conclusions, and limit conclusions that seem implausible. It brings the output of language models closer to human cognitive abilities.

This allows Bridgewater to benefit from the creative scale of large language models while controlling for their limitations. The AIs come up with novel angles no human could conceive of, but humans provide the necessary oversight and skepticism about what may actually play out.

While large language models are far from fully taking over investment functions, Bridgewater’s experiments demonstrate how AI could become a powerful force multiplier for human financial analysts. The future of finance may lie in this type of symbiotic relationship between human and machine intelligence.