How generative AI tools are changing systematic investing

The generative artificial intelligence tools that have captured the public imagination this year have also created significant advances for systematic investing, says Dennis Walsh, global co-head, Quantitative Investment Strategies, Goldman Sachs Asset Management. New large language models that recognize the context of an executive’s comments on an earnings call or the significance of a financial news story open new avenues in the effort to find useful investment signals in unstructured data, he says.

Still, Walsh characterizes the implications of these developments to be more evolutionary in nature rather than revolutionary. “It’s not going to revolutionize our approach to investing, rather, we are excited to leverage these techniques to help us continually enhance how we evaluate investment opportunities,” he says, pointing out that QIS has for decades made extensive use of data, statistics, and econometrics. This has been further supplemented with AI, machine learning, and even natural language processing for more than one decade. New data tools must be complemented by a depth of institutional knowledge — and controls — to develop successful investment strategies, he explains.

We spoke with Walsh to learn more about the history of technology in systematic investing and to hear his views on the latest AI advances.

From where you sit, how big a deal are the new generative AI technologies such as OpenAI’s ChatGPT?

We are seeing a significant breakthrough with the large language models that are based on what is known as “transformer technology” — that’s the T in ChatGPT or in Google’s BERT. This technology introduces contextual relationships between words and documents in a way that is highly efficient and practical, allowing us to train these models using much more data than was computationally possible historically. What has resulted is really a stepwise increase in the power of these models. The jump in efficacy of the latest models relative to prior techniques is meaningful.

How has the technology that QIS uses to analyze data and information changed over the decades?

We have been a systematic investor in equities going back to 1989. For what we do, we need to synthesize information from data that is then used to construct investment signals. If you go back 20 years, that was a relatively simple process. We were generally looking at what would be called structured data, from a balance sheet or income statement or some other traditional sources, going into a database and pulling out a particular line item.

Our investment process and signal research has evolved closely alongside the latest in data and quantitative techniques. Many of the valuable data sets we leverage today are larger, less structured, and generally more complex in nature relative to what was previously available. This also means they require more robust tools and techniques to analyze. Think in terms of financial news articles, earnings call transcripts, analyst research reports, regulatory filings. As technologies progressed over the years, we were able to benefit from the exponential growth of data and start using more unstructured data. 

At first, how we would represent textual data was relatively simplistic. We would analyze a document simply based on the words that were present in the text and their corresponding frequencies. While this enabled us to identify what were the key topics or themes within a body of text, it was a simplified understanding that really didn’t capture the inherent meaning or context within the text. More recently, in the 2015 or 2016 timeframe, models began to be developed that did add some semantic meaning behind words.

With the latest leap in AI technology, you can work with meaning and context. How does that change things?

Context is so important, particularly in our business. To form the view around a company, we may be looking for subtle clues from management or an analyst about what they feel about the true prospects of that company. Having these tools now that can help us process contextual information has led to significant improvements in our ability to extract some of the nuance.

To give you an example that I think is very intuitive from an economic standpoint but also valuable from a security pricing standpoint, we may want to look at how the management of a company views its future growth prospects, compared to what the market believes. For many years now, we have researched and implemented signals that seek to capture whether management is exhibiting bullish or bearish sentiment. We would look at what management would say when they host an earnings call, particularly during the question-and-answer part where comments tend to be less scripted and thus expose more human biases and tendencies. That’s one place where we can really make use of these new tools to extract greater insight from these subtle nuances.

As companies rush to explore the new AI technologies, are you concerned about shortages of the powerful chips needed to run these models?

One of our biggest competitive advantages as an investor is being part of this firm and having access to its resources. The number of CPU cores and graphics processing units we’re using has increased exponentially as we have moved to make use of these large language models. As part of Goldman Sachs, we have access to an infrastructure that combines the best of what we have on-premises with access to the cloud. This has given us the elastic compute capability we need when we’re training a large model. That’s incredibly important.

One of the fears is that AI is coming for our jobs. How do you see AI affecting what you and your researchers do?

We very much see it as a productivity boost — a tool that will help our people perform their research much faster. As an example, we are looking at creating a tool that would understand our internal programming language and our systems and could sit alongside our researchers. The idea is that the researchers could describe in a few words of text the type of strategy they wish to design, and the tool would create the code to do it. That would save hours or days. I think there’s reason to be very optimistic about these technologies.

This article is being provided for educational purposes only. The information contained in this article does not constitute a recommendation from any Goldman Sachs entity to the recipient, and Goldman Sachs is not providing any financial, economic, legal, investment, accounting, or tax advice through this article or to its recipient. Neither Goldman Sachs nor any of its affiliates makes any representation or warranty, express or implied, as to the accuracy or completeness of the statements or any information contained in this article and any liability therefore (including in respect of direct, indirect, or consequential loss or damage) is expressly disclaimed.