Abstract
Updating observations of a signal due to the delays in the measurementprocess is a common problem in signal processing, with prominent examples in awide range of fields. An important example of this problem is the nowcasting ofCOVID-19 mortality: given a stream of reported counts of daily deaths, can wecorrect for the delays in reporting to paint an accurate picture of thepresent, with uncertainty? Without this correction, raw data will often misleadby suggesting an improving situation. We present a flexible approach using alatent Gaussian process that is capable of describing the changingauto-correlation structure present in the reporting time-delay surface. Thisapproach also yields robust estimates of uncertainty for the estimatednowcasted numbers of deaths. We test assumptions in model specification such asthe choice of kernel or hyper priors, and evaluate model performance on achallenging real dataset from Brazil. Our experiments show that Gaussianprocess nowcasting performs favourably against both comparable methods, and asmall sample of expert human predictions. Our approach has substantialpractical utility in disease modelling -- by applying our approach to COVID-19mortality data from Brazil, where reporting delays are large, we can makeinformative predictions on important epidemiological quantities such as thecurrent effective reproduction number.