AI allows investors to process vastly more data. But if investors increasingly use similar models to interpret that information, their conclusions may also be similar.
David Trainer, CEO of the AI-based research firm New Constructs says this is distorting markets and “undermining what we need for wisdom of the crowds to really work, which is diversification of opinion.” When investors’ those views are no longer independent and become the same, “that's what causes manias and bubbles.”
This is nothing new: Investors have been susceptible to groupthink and herd mentality long before AI. In fact, Trainer started New Constructs to help battle that. “Everybody thought the wrong thing about tech stocks during the internet bubble. Everybody thought the wrong thing about housing stocks,” he said.
And AI as a research tool is accentuating that. “When you ask AI, ‘what are the best stocks?’ it's going to go with what's most popular, and popular is not always good,” Trainer explained.
New Constructs uses proprietary machine-learning technology to analyze financial statements and filings to determine a company's underlying profitability and valuation. According to its website, “the firm shines light into the dark corners (e.g. footnotes) of millions of financial filings and provides superior investment research.”
The National Bureau of Economic Research found that leading large language models allocated an average of 41 percent of their model portfolios to semiconductor stocks and favored companies receiving nearly 10 times more news coverage than the average public company.
“AI recommendations are primarily driven by the media attention that firms receive," the NBER report stated, noting that while the portfolios beat the S&P 500, the returns aren’t abnormal when factoring trading costs.
Trainer doesn’t always trust the information coming out of a handful of companies that dominate AI today. While an extreme view, Trainer believes incidents in which OpenAI and Anthropic disclosed that pilot versions of models escaped a secure testing environment and hacked into other companies to find the answers to the ‘test,’ were publicity stunts for a specific purpose.
“I think Anthropic and OpenAI are purposely staging these escapes to give this impression to the general public that they are so powerful that they have a mind of their own,” he said. Paradoxically, if people believe bots can engineer their escapes and break into other websites, “well then you can believe they can tell you how to invest.”
Meanwhile, companies, and their investors, engage in magical thinking and want to use AI to automate much of their work and cut costs. And Wall Street wants trillion-dollar IPOs. “It's a recipe for disaster.”
When Trainer cut his teeth on Wall Street building manual Excel models at Credit Suisse during the 1990s, he was among a small group of people at the firm “that weren't super bullish about the internet.”
“I had a few friends on the sales desk that were big supporters and loved it,” he said. “But eventually, each of them knew in their hearts, the numbers didn’t add up.”
After the dot-com bubble burst, Trainer realized that if the top global investment bank wasn’t going to create the systems to do painstaking fundamental research, including reading thse complex filings and footnotes, no one was (Credit Suisse played a role in fueling the tech IPO bubble under banker Frank Quattrone).
“It was clear to me that that equity research analysts were no longer incentivized to do that kind of work,” he said. So, that was where the seed for his AI-based research firm was planted.
