In the Beginning Was the Word: LLM-VaR and LLM-ES

Abstract

This study introduces LLM-VaR and LLM-ES, novel approaches utilizing general-purpose large language models (LLMs) for zero-shot forecasting of Value at Risk (VaR) and Expected Shortfall (ES). Using the LLMTime framework, these methods process financial time series data encoded as numerical strings, providing a flexible, assumption-free alternative to traditional risk estimation models such as GARCH and EWMA. Our empirical analysis reveals that LLMs perform effectively within a short-term historical context, particularly in highly volatile markets like cryptocurrencies. However, as the historical context lengthens, the accuracy of LLM-based methods diminishes, with conventional models proving superior for capturing long-term dependencies. These findings highlight the potential of LLMs as adaptable tools for risk assessment over recent historical windows, while underscoring the continued importance of traditional models for robust, long-term financial risk management.

Keywords: Value at Risk, Expected Shortfall, GPT, LLM-VaR, LLM-ES, Large Language Models