Home Publications · Oden Institute

Assessment of Predictability of a Class of Models of Growth of Coronavirus 19 Cases as an Exercise in Predictive Computational Science

J. Tinsley Oden, Prashant K. Jha, Lianghao Cao, Taemin Heo, Jing Hu, Mathew Hu, Jonathan Kelley, Jaime D. Mora Paz, Cyrus Neary, Akhil Potla, Sheroze Sherriffdeen, Chase Tessmer, Christine Yang · Oden Institute Report 20-10 · 2020

Abstract

A report from the Spring 2020 graduate course *Foundations of Predictive Computational Science* at the Oden Institute, taught by J. Tinsley Oden with Prashant K. Jha. The class was set a single question as its final assignment: take a simple growth model of disease cases, and establish how far it can actually be trusted to predict. Each solution calibrates the model against roughly forty days of reported COVID-19 case data from the United States, Japan and South Korea, validates it against a later window, and predicts a quantity of interest — the total case count at around one hundred days — with its uncertainty. The solutions differ in the metrics and validation tolerances they choose, which is part of the point. The report's own conclusion is the honest one: the spread of predicted quantities suggests the simple model **cannot** capture the real trend. What the exercise demonstrates is not a forecast but the procedure — how to establish whether a prediction is worth anything before relying on it.

Written up from the class solutions, and a compact demonstration of the argument that runs through this whole program: a model’s usefulness is a claim that has to be established, not assumed.