Application of Large Language Models in Financial Markets
OpenAI
Hudson River Trading
Quantitative traders are exploring whether Large Language Models (LLMs), successful in natural language, can predict financial time series. Challenges include high noise-to-signal ratio, less structured data, and active competition. However, techniques like multimodal learning, residualization, and synthetic data generation show promise.
The article examines the potential of Large Language Models (LLMs) for financial market prediction. It notes that LLMs are autoregressive learners, similar to models used in quantitative trading. However, financial data is noisier and less structured than language, and market efficiency makes prediction difficult. At NeurIPS 2023, Hudson River Trading compared GPT-3's training tokens (500 billion) to annual market data tokens (177 billion for US stocks). Multimodal learning could combine price data with alternative data like news or satellite images. Residualization in finance (factor models) and AI (residual networks) share a similar idea of predicting deviations from a common structure. LLMs' long context windows can capture multi-scale market phenomena, but predicting multiple future time horizons remains challenging. Synthetic data generation could augment scarce market data, and generative models might help simulate extreme events. The article concludes that while LLMs are unlikely to take over quantitative trading soon, they may boost fundamental analysis and urges keeping an open mind.
- Сокращения
- LLM = Large Language Model — Большая языковая модель
Source: The Gradient —
original