search

epiforecasts / covid.german.forecasts

Comparing crowd sourced and model derived forecasts of Covid-19 for Germany and Poland
https://cmmid-lshtm.shinyapps.io/crowd-forecast/
MIT License
4 stars 1 forks source link

Adds deaths predicted by cases convolution model #28

Closed seabbs closed 3 years ago

seabbs commented 3 years ago

This PR adds the deaths predicted by cases model recently implemented into the development version of EpiNow2 along with a test forecast for this week for checking purposes. The model is purposely simple in that deaths are entirely based on cases along with a delay from case report to death and a scaling of cases reported to cases that ultimately die.

As implemented it can only be run when rt-forecast/update-case.R has been run for the target date as this generates the required predictor.

This weeks retrospective submission (as a test) looks within bounds of acceptable without being staggeringly convincing. The breakdown in Germany appears to be due to the model struggling with the last week of data compared to previous data. Hard to say if this is a fault of the model or something actually present in the data without formal evaluation.

Screenshot 2020-12-02 at 16 18 21 Screenshot 2020-12-02 at 16 18 37

Fit to data

Germany-fit Germany

Poland-fit Poland

FYI: Checked with Johannes and all good to submit but may not be included in the visualisations etc. for now.

nikosbosse commented 3 years ago

looks good! haven't checked out the forecast_secondary function as I couldn't find that. Only thing that needs to be addressed I think is the naming + folder of the submission file

seabbs commented 3 years ago

The bit that is missing here is that it assumes that observed cases (i.e with a day of the week effect) are true. It might be a better bet to use inferred infections or day of the week adjusted reports.