How Effective Are Time Series Foundation Models for Epidemiological Data Analysis?

Lucas Moreira Ribeiro, Eduardo Luz, Jadson Castro Gertrudes

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Time Series Foundation Models, Public Health, Epidemiology, Evaluation models

TL;DR: This study evaluates Time Series Foundation Models for zero-shot epidemiological forecasting against classic models. It investigates TSFM architecture's influence, effects of forecast horizon/context length, and seasonality capture without timestamps

Abstract: Understanding the capabilities and limitations of modern time series models is essential for their effective use in domain-specific contexts. This study investigates the zero-shot forecasting performance of state-of-the-art Time Series Foundation Models (TSFMs) using real-world epidemiological data. Guided by a central hypothesis and three research questions, we benchmark several TSFMs, including TimesFM, ChronosT5Base, and Moirai variants, on monthly time series of notifiable diseases in Brazil, considering different forecast horizons and context lengths. Through robust evaluation using MASE and CRPS metrics, along with statistical significance testing, we find that TimesFM and ChronosT5Base consistently outperform classical statistical baselines, while others demonstrate architectural limitations. These findings contribute to a deeper methodological understanding of TSFMs in health-related forecasting scenarios and suggest that performance differences may arise from model design and training data characteristics. The results provide valuable insights to guide future model development and support more informed use of forecasting tools in public health decision-making.

Track: 5. Public Health Informatics

Registration Id: X9NR8X79Q3X

Submission Number: 18