LLMs are the new pied pipers of the stock market

**Rachel Varghese** (0:01)
If you've ever asked AI to pick your next stock investment, you might want to think again. Let me explain why. This June, HSBC released a survey of 1,000 rich Indians, in which it found that more than three-fourths of them actively use AI for their stock market investments. Out of them, more than 40 percent even went on to admit that the AI tools they were using were the main source of their ideas.
But think about which AI chatbot is always open on your laptop. Chances are, it's most likely either OpenAI's ChatGPT, Anthropics' Claude or Google's Gemini, which raises a question my colleague, Mutasim Khan asks in his piece. If everyone is using the same three chatbots for their stock picks, what goes wrong? To answer that question, I'm going to be reading out an edition from one of Ken's most popular subscriber-only newsletter called Kacheng, and it's titled, LLMs are the new Pied Pipers of the stock market.
Welcome to Daybreak, a business podcast from The Ken. I'm your host, Rachel Varghese, and every day of the week, my co-host, Snigdha Sharma and I will bring you one new story that is worth understanding and worth your time. Today is Monday, the 17th of August.
I spent a summer as an intern on a trading desk once, and realized fairly quickly that I was not very good at it. This was long enough ago that good old public forums like Stack Overflow and R slash Learn Python were still the go-to over ChatGPT for coding advice. The fun part of my job was making wild hypotheses about the market and hoping they work. All the rest, like writing code, fetching data, learning new strategies, made me yawn around a little too much for my manager's liking. But the same friction that deterred me from turning an idea into an actual trade is also the friction large language models for LLMs have been easing remarkably fast ever since. An HSBC survey of about 1000 affluent Indian investors released in June found that 86% of the respondents are now using AI in their trading flows, which is the highest share of any country the survey covered, about China, Singapore and the US. Forty-two percent of them counted AI-powered tools among their leading sources of investment ideas, second only to China's 48%.
But individual investors are not the only invitees to the LLM gala. Finfluencers or social media content creators who share financial advice, are stacking ChatGPT prompt packs on tops of the tips, courses and telegram boards that they are already selling. The internet is also flooded with guides like 50 ChatGPT prompts for intraday and short-term traders, and entire prompt libraries, almost all of them opening with some variant of act like you are a professional Indian market analyst.
If prompts feel like too much effort, GitHub, which is Microsoft's developer platform, is full of open-source trading boards that plug an LLM into your brokerage account, allowing a simple chat with say, Claude to drive your trades.
It appears that LLMs are becoming more and more embedded into markets, from every direction.
The premise behind this whole movement is essentially that AI is on its way to leveling the playing field between individual and institutional traders, and that early movers get a short but glorious window of easy money before the rest of the market catches up and closes it. But still, there is a catch. LLM usage is almost entirely concentrated within three models. OpenAI's ChatGBT, Google's Gemini and Anthropics Cloud. And a growing body of academic work has begun documenting something specific about what these models tend to do at scale.
LLMs by design are non-deterministic. No two answers are ever identical even to the same prompt. But run the same kind of query enough times across a large enough sample and a distribution emerges. When it comes to financial markets, LLMs have been shown to have systemic biases in that distribution. Biases not toward any particular stock per se, but towards stocks with certain characteristics. They skew, for example, toward large-cap stocks over small ones, toward technology over other sectors, toward contrarian setups over momentum, and toward buying rather than selling. The paper argues that since major LLMs are trained on broadly the same universe of public financial writing, they end up sharing the same systematic biases in the aggregate.
Which is a strange kind of level playing field. The retail trader gets tools that, in the aggregate, make her more predictable to the very institutions she thought she was catching up to. In a June podcast, Osman Ali, the global co-head of Goldman Sachs' quantitative investing arm, put it plainly. He said, if you ask the same model the same kind of question, you get the same type of answer. Which means, investors pile into the same type of securities at roughly the same time and prices get pushed away from any sort of fundamental value. Whether or not retail investors were consciously using LLMs to make their decisions, Ali said that the crowding effect was already visible. Broad LLM use had produced a different type of predictability and inefficiency in the market.