This algorithm was developed using the now published paper by Timothy W Russell et al. The publication was at the time a preprint and the algorithm follows the state of the code as in july 2020. The modifications made were of language and deployment. The original code was translated from R to Python to made use of functionalities like tensorflow 2, GPU parallelization and others. The number of chains drawn was also reduced in order to operate in time for the release of the nowcasting report everytime the Health autorities publish data.
For the reasons stated previously, to perform a more robust analysis, larger chains and more chains must be sampled and tested to converge.