ATD: Dynamic Modeling for Extreme Event Prediction with Uncertainty Quantification with Multi-panel Time Series

The ability to predict extreme geopolitical events is crucial for national security and foreign policy decision-making in a rapidly evolving world. A reliable and effective prediction system for this purpose must also provide a quantification of prediction uncertainty to support informed decision-making. In this project, the investigators aim to construct an advanced and comprehensive prediction system based on statistical models that interpret the dynamic relationship between multiple event series at multiple locations. The sophisticated statistical prediction system with quantified uncertainty has numerous real-world applications in various threat detection and risk assessment scenarios, such as cybersecurity, network activity monitoring, weather pattern forecasts, epidemics tracking, and others. The ability to effectively predict and assess potential threats will help organizations make informed decisions to ensure safety and security. This project will provide early-career students with valuable, interdisciplinary research experiences, offering a unique opportunity for growth and development. The investigators are committed to promoting diversity and inclusion in STEM fields, and will actively seek to recruit students from groups that are historically under-represented in science and engineering. A comprehensive project website will be created for project papers, reports, presentations, and links to relevant resources, providing a centralized repository of information and resources. The project’s outcomes, including the developed methods and software, will be widely disseminated for public use.The investigators will develop a dynamic matrix factor model for count series, which leverages a generalized linear factor model to reduce dimensionality and the dependence of the latent factor process to make forecasts. They also will develop a generalized matrix autoregressive model specifically tailored to low dimensional series (e.g., regional special events). The recently introduced repro-samples approach for irregular inference problems are used to assess model uncertainty. Given the complexity of the data, combining different models that better predict different parts of the data is beneficial. Two model averaging schemes are used by combining predictive distributions and dependent p-values, both of which come with uncertainty quantification. This leads to a warning system that predicts the likelihood of an extreme event occurring, as measured by a confidence level. The overall framework and methodology can be easily applied to other fields requiring prediction-based surveillance and monitoring. The resulting framework and methodology will have a profound impact on other fields of statistics, including high-dimensional statistics, statistical analysis of matrix and tensor data, modeling large panels of dependent data, statistical inference of irregular problems, predictive inference, fusion learning, and more. It will significantly advance statistics and data science research in general and provide new insights into data-driven decision making.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.