The article discusses how Large Language Models (LLMs), which power products like ChatGPT, are being applied in financial markets, driving significant investment and interest. These models excel at modeling sequences of tokens, making them valuable for financial data analysis.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Quantitative traders are attempting to shift LLMs from text prediction to price prediction, but the extremely low signal-to-noise ratio of financial data and adversarial competition make this path fraught with challenges.
The core question posed by quantitative traders is whether LLMs can be shifted from modeling word sequences to modeling price or transaction sequences. As autoregressive learners, LLMs use previous elements in a sequence to predict the next element, which aligns with the research direction in quantitative trading of finding autoregressive structures (e.g., statistical arbitrage). However, the quantity and information content of financial data become key bottlenecks.
At the NeurIPS 2023 conference, Hudson River Trading compared GPT-3's training token count with the data volume of the U.S. stock market: assuming 3,000 tradable stocks, 10 data points per stock per day, 252 trading days per year, and 23,400 seconds per trading day, there are approximately 177 billion market data tokens per year. GPT-3 was trained on 500 billion tokens, so the orders of magnitude are similar. But the key difference is that financial tokens are prices, returns, or volumes, not words or syllables, and their predictability is far lower than that of language.
Language has an underlying grammatical structure; humans can easily predict the next word in a sentence, but can hardly predict the price change of the next trade. The market is filled with many smart people competing, and any signal is quickly arbitraged away, making the market 'efficiently inefficient.' Moreover, financial data contains far more noise than signal: the trading motives of individuals and institutions can be irrational, and the market is constantly affected by new information, regulatory changes, and macroeconomic shocks. In contrast, language evolves slowly and stably.
Multimodal learning is an emerging direction in AI research, potentially integrating different modalities such as prices, news, social media sentiment, and satellite imagery to improve predictive power. The idea of residualization is applied in both finance and AI: financial factor models decompose common components and idiosyncratic components, while AI residual networks learn residuals from identity mappings; both aim to learn predictions more efficiently. LLMs' long context window capability can capture multi-scale phenomena in financial markets, from microsecond-level order book imbalances to multi-month fundamental information.
Current Transformer models only predict the next time step, but financial trading often requires predicting the complete term structure of returns over multiple future time steps. Synthetic data generation is another potential application of LLMs: by simulating price trajectories, it can alleviate the scarcity of real data and support meta-learning techniques—first training strategies in low-cost simulators, then calibrating with real data, similar to approaches in robotics.
This article is based on a report from The Gradient, which cites data from Hudson River Trading's presentation at NeurIPS 2023 and multiple academic papers. All figures and conclusions are derived from the original source, without introducing external knowledge.
LLMs face fundamental challenges in financial prediction such as low signal-to-noise ratio and adversarial competition, but techniques like multimodal learning, residualization, and long context windows still hold potential; synthetic data generation may be a more pragmatic application direction.
Primary report
Primary source