Financial Market Applications of LLMs
The Gradient Richard Dewey
Quant researchers ask if ChatGPT-style models can predict stock prices instead of words. Turns out markets are way noisier than language, so don't expect an AI trading takeover yet.
Based on reporting by The Gradient, Richard Dewey — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Every AI hype cycle eventually hits Wall Street, and this one is no exception. A new piece from Richard Dewey and Ciamac Moallemi at The Gradient digs into whether the same transformer architecture powering ChatGPT could be repurposed to predict prices and trades instead of words. The pitch sounds obvious: LLMs are autoregressive, and so is most quant trading. Just swap tokens for ticks.
The data math is more encouraging than you'd expect. Hudson River Trading crunched the numbers at NeurIPS 2023 and found the US stock market generates roughly 177 billion tradable data points a year, based on 3,000 stocks, 10 observations per stock per day, and 252 trading days. GPT-3 trained on 500 billion tokens. Not the same ballpark, but not absurdly far off either.
The problem isn't volume, it's signal. Language has grammar, structure, and authors who want to be understood, which makes the next word fairly guessable. Markets have none of that. Prices are shaped by thousands of smart, adversarial participants actively erasing any exploitable pattern, plus a healthy dose of irrational noise, like the 2021 GameStop saga, that has nothing to do with fundamentals. Economist Lasse Pedersen's phrase
My take — AI-written commentary, not fact-checked reporting
efficiently inefficient" captures it well: markets fight back against predictability in a way sentences never do. Still, the authors don't dismiss AI's usefulness in finance, they just redirect it. Multimodal models could fuse price data with satellite images of shipping ports or Twitter sentiment. Long-context attention could help spot patterns across wildly different timescales, from second-level order book noise to month-long earnings drift. And synthetic data generation, already used in robotics to pretrain controllers cheaply before fine-tuning on real hardware, could let trading strategies bootstrap on simulated markets before touching real capital. The most plausible near-term win might be humbler: LLMs as tireless research assistants, flagging contradictions in earnings calls or surfacing links between unrelated industries, essentially a Charlie Munger on demand rather than a rogue trading bot. The honest conclusion is that nobody predicted transformers would get this good at language, so nobody should be too confident about ruling out their next trick either. The authors aren't betting on GPT-5 replacing quant desks. But they're not betting against surprise, and in this field, surprise has had a pretty good track record lately. <opinion>I've watched enough "AI will crack the market" cycles to know the pattern: the hard part was never architecture, it's that price data is adversarial while language mostly isn't. The genuinely useful application here isn't LLMs trading for you, it's LLMs doing the unglamorous grunt work of cross-checking earnings calls faster than any analyst, and anyone pitching more than that is selling hype, not alpha.
Read more about this at: The Gradient