A technology-first approach to investing money is not new, but the tools in artificial intelligence are giving the business new opportunities to outperform.
Jeff Shen, San Francisco-based co-chief investment officer and co-head of systematic active equity, says language-learning models are becoming powerful tools.
“We’re in the midst of a revolution,” he said. “Big data, alternative data, and now generative AI is transforming all industries, including asset management. There’s more data available and better algorithms to capture that data, and that makes systematic investment exciting.”
Four decades of quant
The systematic team’s origins are the Barclays Global Investors business that BlackRock acquired in 2009. The deal emerged when Barclays, hit hard by the global financial crisis, surrendered its investment business to survive – and made BlackRock the world’s largest asset manager, then at $2.7 trillion.
BGI’s roots go back to 1985 as what today might be considered a fintech: a Silicon Valley-based operation using big data and primitive forms of machine learning, long before these terms or capabilities came into fashion. It is a quant shop, using data-driven insights to zero in on lots of small, rapid bets arbitraging one stock against another – Coke versus Pepsi.
This works even if the industry or market is doing badly – Country Garden versus Evergrande. What counts is finding a tiny, short-lived edge that the manager can trade quickly, at scale, and then close the position. Multiply such trades by hundreds or thousands across a portfolio, and the firm creates a large equity strategy with a low correlation to benchmarks.
With more data, better algorithms, increasing computing power, and the electronification of stock markets, BGI had emerged as a cutting-edge powerhouse and continues as BlackRock’s systematic arm.
Since then, the ETF world has taken off, making BlackRock the world’s biggest asset manager. As of September 2023, the firm reported $3.1 trillion in exchange-traded funds (a retail business) and another $2.6 trillion of index funds (for institutions). The firm’s technology services group, including its Aladdin portfolio risk system, is another important contributor to revenue.
AI’s progress
In this context, the systematic equities business, an institutional business, is modest, at $237 billion of assets under management. Shen is of course bullish about his division. “Systematic quant investment is now in a golden age,” he said.
But excitement around generative AI, which includes natural-language models such as ChatGPT, gives Chen’s optimism some credence.
In the old days, quant tactics consisted of ranking US large-cap stocks by traditional metrics (price to book, price to earnings, dividend yields). Even then, the biggest quant hedge funds built data warehouses of astonishing size. This gave them the ability to generate performance regardless of market trends. The most successful firms made a lot of money, led by Renaissance Technologies, which from 1988 to 2018 was the world’s most profitable (and secretive) investment firm.
The steps involved in running active strategies, quant or otherwise, have steadily automated. Information is now machine readable, such as broker reports, company financials, media stories, and government statistics. Natural-language processing made it possible to turn unstructured data (anything from a PDF to a lawyer’s signature) machine-readable. The Internet of Things and satellite imagery have expanded the list of things that can be measured and quantified. Moreover, these now give fund managers access to real-time views.
Shen cites the movement of trucks. Geospatial tagging, WiFi beacons and satellite images enable buyers of this data to track fleets of trucks. This gives them a sense of traffic between suppliers and stores, one data point to determine how a company is faring. Build enough of these, and a firm can widen its scope to get a macro view of the economy.
Enter GenAI
Today generative AI is adding a new set of tools to the mix. But it’s not just another way to crunch data. It actually changes the way portfolio managers understand information.
Shen gives the example of a news report about a CEO stepping down. For the past twenty years, tech-savvy firms used machine learning to follow a ‘bag of words’ approach. The machine would parse a text and look for concentrations of words or phrases that correlate to good or bad, buy or sell.
In the example of the CEO losing his job, the machine might identify seven relevant wordings in the opening paragraph. It would tag as negative clusters such as ‘alert’, ‘leaving the company’, ‘replaced’, ‘frustration’, and ‘weaker’. It would also highlight two upbeat expressions, ‘surprising’ and ‘respond positively’, but overall the weight of negativity would lead the computer to recommend a sell.
If this company were part of a Coke versus Pepsi duo, BlackRock might decide this was a signal to go short one and long the other, with leverage. The trade might last a few hours or a few days, but the speed of the analysis would give the team a different outcome than the mass of active fundamental players relying on a human interpretation.
“That was the state of the art in 2007,” Shen said. Since then, the data and the aglos have gotten better, but the bag-of-words approach was still the norm. LLMs such as ChatGPT are changing this.
LLMs take the same paragraph and, in Shen’s example, conclude it is a massive positive rather than a piece of bad news. That’s because it’s not just translating text, but understanding it in context. The LLM knows that, while there’s a bunch of negative words up top, the key phrase is at the bottom: ‘we expect the stock to respond positively’.
“Despite this being news about a CEO stepping down, the LLM understands gist of the press release – it gets the punchline,” Shen said.
Data and algos
Although this example is designed for BlackRock presentations to journalists, the implication is that a systematic shop adding LLMs to the mix should perform better. In this tidy example, in fact, the portfolio manager is given a completely different answer.
Real life isn’t that neat, but Shen says LLMs are the next wave of tools designed to give a manager a tiny edge. Firms such as BlackRock are now using LLMs on proprietary data sets, in order to train the models on financial and other specific types of data. He says BlackRock finds its proprietary LLMs have an edge over ChatGPT (which is trained on the internet at large).
This brings quants back to the same old basics: who’s got the best data and the best means of scrubbing it; and then who’s got the cleverest algos. But LLMs add another wrinkle here as well, by helping humans improve their means of judgement.
The human touch
Although some quant shops such as RenTec were notorious for just following their computers, Shen says systematic strategies still require human decisions. This becomes clear at times when the historic data is incomplete or doesn’t’ exist. For example, modeling a company during Covid was hard because the last global pandemic of this magnitude occurred a century ago. There’s no reliable data from 1918 to use today. So while quants used real-time data around traffic or job postings to get a view, it still required a human to extrapolate what this meant for the near future. Big data, on its own, was not a reliable predictor.
But with LLMs, the humans can ask the machine nuanced questions that were impossible to ask a machine-learning system. This turns the LLM into a productivity tool, and different questions lead to different outcomes. The old big-data models of the 1980s and 1990s rested on parsing valuations, and in the 2010s added things like market sentiment. Now the scope of questioning is broad, which enables human creativity.
“The question can be a competitive edge,” Shen said.
Given what Shen depicts as a bright future, does this suggest active management styles will begin to outperform versus passive strategies? Are systematic investments poised to claw back some of the assets that have flowed to the ETF side of the house?
Shen remained non-committal. The industry winners, he says, are those firms that embrace AI, regardless of the product. A safe answer. Therefore, a safe supposition will be that the new competition using tech will advantage firms with the resources to get their hands on as much data as possible.
