Zero-shot forecasting of epidemics

Accurate epidemic forecasting is critical for mitigating the global burden of infectious diseases, enabling timely interventions and optimal resource allocation. While conventional forecasting approaches can effectively model the temporal trajectory of epidemic dynamics, they rely on task-specific learning from historical observations, which are often scarce and limited. Recently, pre-trained Large Language Models (LLMs) have emerged as powerful foundation tools for zero-shot time series forecasting across diverse domains, eliminating the need for retraining on task-specific data. In this study, we conduct a comprehensive evaluation of LLM-based foundation models to capture the complex dynamics of epidemic incidences across multiple temporal horizons. Our analysis not only benchmarks these models against statistical and deep learning frameworks, but also identifies which architectural designs of LLM models are suitable for epidemic forecasting. The empirical results conducted on eleven epidemic datasets spanning three diseases from distinct geographical locations highlight that TimeGPT and TiRex models exhibit superior generalization capabilities. These findings underscore the potential of zero-shot LLM-based epidemic forecasting to support effective decision-making.