Forecast Intervals for Infectious Disease Models

Forecast intervals for infectious disease transmission and mortality have long been overconfident — i.e., the advertised coverage probabilities of those intervals fell short of their subsequent performances. Further, there was no apparent relation between how good models claimed to be (as measured by their purported forecast uncertainties) and how good the models really were (as measured by their actual forecast errors). The main cause of this problem lies in the misapplication of textbook methods for uncertainty quantification. A solution lies in the creative use of predictive tail probabilities to obtain valid interval coverages. This approach is compatible with all probabilistic predictive models whose forecast error behavior does not change “too quickly” over time.