Large Language Models (LLMs) excel at predicting the next token in a sequence. Quantitative traders are investigating whether this ability extends to predicting sequences of prices or trades rather than words.
At the 2023 NeurIPS conference, Hudson River Trading presented data comparing LLM training volumes to market data availability. With 3,000 tradable stocks and 10 data points per stock daily, there are approximately 177 billion stock market tokens generated annually. This volume is comparable to GPT-3's 500 billion token training set.
However, predicting financial returns presents significant challenges compared to language. Markets are highly efficient due to competition among smart traders, leaving little signal. Additionally, financial data contains more noise than signal, driven by non-rational trading and frequent regulatory or macroeconomic shifts.
Multimodal learning offers a potential avenue for finance. Models like OpenAI's DALL-E 2 can process both text and images. In finance, this could combine technical time series with alternative data such as social media sentiment, news articles, and satellite imagery of shipping activity.
The concept of residualization appears in both fields. Financial factor models remove shared market components to focus on idiosyncratic asset returns. Similarly, transformer architectures often learn the residual of a function rather than the function itself to improve efficiency.
Source: https://thegradient.pub/financial-market-applications-of-llms/
