I’m currently working on nowcasting COVID-19 cases by specimen date (date at which test was taken) as part of my MSc project. The data available here presents a really interesting opportunity to explore the impact of reporting structures on the relative performance of nowcasts. If anyone is interested, the code for my project is available in this GitHub repo.
In particular, I’ll be looking at how accounting for various reporting effects (e.g. no reporting on weekends and public holidays, reporting rate varying by day of week, differing reporting frequency) can affect the performance of nowcasts as well as the model diagnostics and runtime. My hypothesis so far is that more complex models generally perform better but at the cost of increased runtimes and likelihood of running into diagnostic issues.
Looking forward to sharing more as the project goes by!