(Bloomberg) -- Jesse Livermore scanned trendlines. Warren Buffett sought a margin of safety. Peter Lynch bet on growth rates. In the long history of markets, trading systems and investment formulas hold an honored place.
Yet even the finance legends couldn’t have predicted what artificial-intelligence proponents are dreaming up these days, thanks to the computational firepower of language models like ChatGPT.
In this new world, there are automated programs that blare out warnings if corporate executives go on unnecessary tangents or jump between subjects — potential signs of anxiety about the future. Another AI model dissects product blueprints and graphs from business slides in an attempt to forecast stock swings. There’s even an trading tool that compares actual statements from the C-suite with imaginary dialogue cooked up by machines to figure out — somehow — market liquidity.
The list of ingenious-sounding investing ideas churned out by academia is growing by the day.
Of course, quants have spent decades trying to uncover hidden stock omens across Corporate America with mixed success. But natural language processing — the branch of AI that deals with text comprehension — is the hottest toy again thanks to the wonders of chatbots. The rush to cash in is leveraging long-standing ties between university researchers and systematic investors — and opening a new frontier in so-called sentiment analysis.
“Basic sentiment analysis based on dictionary — absolutely, that is being arbitraged away,” said Mike Chen, the head of alternative alpha research at Robeco, which runs $82 billion in quant strategies. “There’s so much more you can do.”
At its core, linguistic data-crunching seeks to help quants get better at predicting the future by analyzing the meaning of the text behind the numbers. Think of an analyst reading the news or listening to earnings calls, but scaled up across a myriad number of sources tracking every single company at every single moment — in the blink of an eye.
Beyond text analysis, there’s a plethora of research dissecting speaking tones, facial expressions in videos and even emojis. At Robeco, quants recently started analyzing the tone and pitch of executives on earnings calls for a sign of their true confidence. Stockpulse, a data provider, this year began compiling what influencers are saying about the economy and various companies on TikTok.
Over at AllianceBernstein, data scientists Andrew Chin and Yuyu Fan are going all-in on AI tools to find hidden meanings in corporate spiel. Not every attempt has worked. For instance, when they dug into how Chinese companies summarize on-site visits from brokers, they found that the more complex the text — like the length of sentences and unnecessary words — the more evidence that the firm in question is struggling.
On the other hand, the number of words divided by the length of the broker event, a proxy for speaking speed, didn’t mean much. In US earnings calls, they studied the use of the “we” pronoun as a sign of collaboration and unity. That also proved meaningless.
“We really try to generate a broad set of signals — sometimes hundreds — but it doesn’t mean all of them will work,” said Fan, a senior data scientist at the $669 billion manager.
Whereas machine reading used to rely on counting positive and negative words, the large language models behind chatbots are far better at parsing context even across meandering paragraphs. These robots not only purport to structure the unstructured, they also promise to automate research tasks and generate fresh trading ideas at breakneck speed. And the introduction of ChatGPT alone — which has ingested enough text that it has a good grasp of all subjects — is a regime shift. Academic researchers have found that simply telling the chatbot to rate if a news headline is good or bad for a stock has produced better results than prior methods.
At the same time the AI hype also reflects an awkward fact of academic life for quant professionals: Build a new research toy, and the papers will come. Long before ChatGPT, market practitioners were casting a wary eye on the glut of freshly discovered ways to pick winning stocks. There’s even a snide name for it: the “factor zoo.”
New research techniques have sped up the boom in trading ideas, but there’s a lot of noise out there.
“Some of them are useful but probably many are not,” said Yin Luo, a quant analyst at Wolfe Research. “There are way too many academic papers in almost every single field about using ChatGPT in many different ways.”
The hunt for new signals is already proving a cat-and-mouse game as corporate executives try to outsmart the robots. A 2020 academic paper showed management personnel are now deliberately deploying positive-sounding words and avoiding the opposite to notch higher sentiment scores. Now as AI tools get more sophisticated, companies are updating their presentations to incorporate a positive vocal tone and upbeat sentences, according to this year’s update to the aforementioned study.
That’s why quants like Robeco’s Chen are looking for subtle signs of management confidence that are harder to fake. At the Dutch asset manager, machines scan corporate speech to detect the concrete over vague, the spontaneous over the scripted and whether executives are answering analyst questions directly.
“Executives have been coached on saying the right words,” he said. “There are a lot of different types of stuff you can look at to go beyond basic sentiment.”
One way to make the latest language models even better at finance is to get humans to label sentences based on their own expert interpretation of whether they are positive or negative, as AB and Man Group have done. That helps provide additional training for the AI.
At Man, quants have been experimenting with what prompts are best for getting ChatGPT to interpret corporate-speak. Feeding it with a few examples seems to help, says Slavi Marinov, head of machine learning at the systematic unit Man AHL.
To him, the current drawback with a lot of sentiment research is even if the signals are predictive, a lot of the time they simply behave like classic quant factors such as momentum, an investing style that rides the winners on the way up and the losers on the way down.
“People might see that in isolation some signals are useful to predict returns, but they don’t look at whether they are additive to all the things that we already know about,” said Marinov. “So they keep rediscovering the same basic effects.”
In other words, insights derived from sentiment analysis can bring little fresh value to trading pros, who are already scrutinizing everything from recent price moves to analyst forecast revisions to ride market trends.
Like all things AI-related then, the boom in sentiment analysis holds real promise even if there’s a lot of hype out there. Yet despite new computing advances, humans are far from obsolete. If anything, industry specialists have become more relevant to the number crunchers — even if it’s just teaching them where to look.
“If you’re an analyst and you’re going to look for this anyway, then it probably makes sense,” said Chin, the head of investment solutions and sciences at AllianceBernstein. “If you just have lots of data and you’re trying to find some patterns and relationships, some of that will be noise and not useful.”
--With assistance from Shelby Knowles Nikolaides.
©2023 Bloomberg L.P.