Analysis pipeline: steps forward

[SamuelBrand1]

This is a meta-issue enumerating the tasks to be done to get to robust/repeatable analysis, as discussed in #193 .

Task list

• Create a DrWatson project for the pipeline. This requires a bit of thought about what features we want. #196
• Within the project setup an appropriate environment, including loading EpiAware #200
• Design generation of "truth" data script as described in README, along with saving data. #202
• Define the full sequence of different scenario options, e.g. as here. #205
• Construct and test a function for going from scenario specification to running inference. At this point defining the output to save is important. #211
• Blend Dagger.jl checkpointing into the scenario inference loop process, e.g. like here. #219
• Prior predictive modelling for each inference model #227
• Write and test a forecast function which projects inference forwards in time. #231
• Evaluation method on inference output. I'm assuming we outsource this to scoringutils. #247

[SamuelBrand1]

I've edited the list to include a forecast function. My current thinking is that we can avoid just appending missing data on the end of the case/observable count time series:

• The underlying inferred latent process in infections (e.g. $Z_t = \log R_t$) depends on an unconditional white noise process we denote \epsilon_t. This can be extended trivially to augmented_epsilon_t.
• Conditioning on this, and other parameters (e.g. the damp parameters in an AR process, the std of the white noise etc) lets us sample $Z_t$ into the future.
• Conditioning on the sampled $Z_t$ and other parameters (e.g. initial incidence rate etc) lets us project the whole process into a forecast.

[seabbs]

Agree. The thing I want to steer clear of here if possible is being able to do this only in some limited subset of cases vs all cases.

[SamuelBrand1]

Agree. The thing I want to steer clear of here if possible is being able to do this only in some limited subset of cases vs all cases.

This should work for any model built on inferring an underlying white noise process.

[SamuelBrand1]

Agree. The thing I want to steer clear of here if possible is being able to do this only in some limited subset of cases vs all cases.

This should work for any model built on inferring an underlying white noise process.

This has turned out to be the hard limit of the approach too as per TuringLang/Turing.jl#2239