Application of Large Language Models in the Problem of Event Forecasting

Abstract

This paper presents a study of the application of large language models (LLMs) to predict events based on LLM agents—autonomous systems that use LLMs for reasoning, decision making, and interaction with the environment. Various architectures of LLM agents are analyzed: cooperative systems (ChatDev, MetaGPT), multiagent debates (MAD, ChatEval), agents for web tasks (WebAgent, WebVoyager), and simulation agents (Generative Agents, EconAgent). Particular attention is paid to the features of predictive modeling based on LLMs, where classical approaches (regression, time series) are replaced by agent-based modeling and predictive engineering. This article presents the results of an experiment on predicting the outcome of a selected conflict using an LLM agent (Mistral, DeepSeek) and the retrieval-augmented generation (RAG) approach based on data from analytical agencies, opinion leaders, and news sources. The convergence of forecast estimates of polarized sources is revealed and requirements for forecasting systems are formulated: weighting sources according to expert significance, filtering neutral data, and sample balancing. Requirements are put forward for the selection of data assessed by simulation LLM agents.