Financial Market Applications of LLMs

Richard Dewey and Ciamac Moallemi, writing in The Gradient in 2024, examined whether large language models (LLMs) can be turned toward predicting prices or trades rather than sequences of words.

LLMs are autoregressive learners that use previous elements in a sequence to predict the next one, and quantitative trading research, including statistical arbitrage in stocks, likewise seeks autoregressive structure linking news, orders or fundamental changes to future prices. The authors describe a data comparison presented by high frequency trading firm Hudson River Trading at the 2023 NeurIPS conference: assuming 3,000 tradable stocks, 10 data points per stock per day, 252 trading days per year and 23,400 seconds in a trading day, roughly 177 billion stock market tokens per year are available as market data, against the 500 billion tokens used to train GPT-3.

The authors state that price, return or trade tokens are much harder to predict than syllables or words. Language has underlying structure such as grammar, while markets are made nearly efficient by competition, a condition they attribute to economist Lasse Pedersen as "efficiently inefficient." No adversary tries to make sentences harder to predict, but financial data contains more noise than signal, they write, citing the 2021 GameStop episode, and financial time series change constantly with new fundamental information, regulatory changes and macroeconomic shifts such as currency devaluations.

Potential applications they discuss include multimodal learning combining price and trade data with sentiment, news articles, corporate reports or satellite imagery of shipping activity; residualization, compared with residual network architectures such as transformers; long context windows for multi-scale phenomena spanning months, days and seconds; and prediction of the entire term structure of expected returns, since current transformer-style models look only one period ahead.

They also describe synthetic data creation, including simulated price trajectories and meta-learning comparable to robotics, where controllers train in cheap simulators before calibration on real experiments, and sampling from extreme scenarios, which they note is fraught because extreme events occur rarely. Despite skepticism about LLMs in quantitative trading, they suggest the models could aid fundamental analysis by helping analysts refine an investment thesis, uncover inconsistencies in management commentary or find relationships between tangential industries. They conclude that a GPT-4-style takeover of quantitative trading is currently unlikely, but advocate keeping an open mind.