Rolling 7 day average

[isaacmg]

Based on @wwymak comments we would like to try forecasting cases on a rolling weekly average. While this removes a degree of granularity it may ease problems with reporting issues.

Acceptance criteria:

• Experiments logged on primary counties list.
• Pre-processing code for merged into the task-ts repo
• Report detailing finding of using the seven day average.

[isaacmg]

So we have done a number of experiments with this setting. The end result has shown that are models are not generalizing well to out of distribution events. Due to this we will continue to run future experiments with enhanced models in order to explore this further.

[isaacmg]

Update I have now analyzed a couple of the models using rolling 7 day average:

• DA-RNN so far performs a bit better than the transformer on LA County but doesn't do too good on cook county (at least not with our current hyper parameter searches).
• Generic transformer doesn't seem to work well at all in this context. However, we haven't tried extensive transfer and more than basic hyperparam tuning.
  To-do:
• Continue testing DA-RNN with hyper parameters (dropout, different decoder sizes) and transfer learning (need to add the embedding layer into repo for that).
• @kritim13 is also working on probabilistic models we hope to merge soon and test.
• Neural ODE's