In the previous post we have shown that Weibull recovery model works well when trying to reconstruct the recovered cases time series from the confirmed cases time series. In that post we have used random simulation to generate fake recovered cases time series. In this post we will use convolution to get the expected recovered cases time series.

Today we celebrate 10 years since our English "Hello World" post! While it was not the first post on Physics of Risk, (we were writing posts in Lithuanian since 2006), nor it was the first post written in English (a post on Kirman's ants was published on April 11, 2010), it is still an important post, which marks our transition to writing in English.

Fig 1.Stock image from pixabay.

It is interesting to see if we will survive another 10 years.

In the previous post we have estimated mean recovery time in COVID-19 Lithuanian data set. We have tried to generate fake recovery time series assuming that recovery times are exponentially distributed. We have failed. This time we will assume that recovery times are Weibull distributed.

Optimizing is easy, if the problem is small and functions involved are smooth with out local minima or maxima. But what happens when the problem is bigger or not so smooth? In these cases it might be impossible to find the best solution, so we have to be happy about finding "good enough" solution. How do we do it? By using the most metal algorithm in computer science.

This summer one of my colleagues asked me if I know what is the recovery rate from COVID-19. While I have read a lot of papers, some of them claiming to use "realistic" parameters, I still had no clue (as those "realistic" parameters were in some cases wildly different). Yet as we have enough data available, we can estimate recovery rate ourselves.