Applying Large Language Models in Financial Markets
OpenAI
Hudson River Trading
Large language models successfully model word sequences, but attempts to apply them to stock price prediction face fundamental limitations: financial time series contain much noise and little signal, and competition among traders makes markets nearly efficient. Nevertheless, promising directions include multimodal learning, synthetic data, and analysis of extreme events.
The article discusses the application of large language models (LLMs) to quantitative trading tasks, particularly price prediction and trading. The authors note that LLMs are autoregressive learners and can be adapted to model price sequences instead of words. However, a key challenge is the signal-to-noise ratio: in linguistics, text has a grammatical structure that is predictable for humans, while financial series contain many random factors, such as irrational trades and macroeconomic shocks. At the NeurIPS 2023 conference, Hudson River Trading compared the number of tokens for training GPT-3 (500 billion) with available market tokens (177 billion per year), showing that data volumes are comparable but price predictability is much lower. Promising approaches include multimodal learning (combining prices, news, satellite images), residualization to extract idiosyncratic return components, and the use of long context windows to analyze multi-scale phenomena (from microstructure to fundamental factors). Synthetic data can help train strategies with subsequent calibration on real data, and generative models can be used for modeling extreme events. Overall, despite skepticism, the authors encourage openness to potential breakthroughs.
Source: The Gradient —
original