Add plan for all results figures

[seabbs]

As title. As part of this work will need new issues to be created for the desired output figures

[seabbs]

This might be good too address before we start knocking out prototypes

[seabbs]

@SamuelBrand1 I think you are thinking about this now?

[SamuelBrand1]

After some f2f chat with @seabbs, here is my plan for first three figures:

• Fig 1: Overview: This fig aims at presenting the nowcasting and forecasting (e.g. all horizon estimates) at determined inference time point for each scenario with each inference model choice but specialised to well specified generation intervals. Inference time point selection: end of biggest $R_t$ transition point in scenario with closest to middle of whole scenario breaking ties (e.g. smoothly endemic scenario). Plotting details: 4 rows corresponding to 4 main scenarios (e.g. outbreak with measures etc.), 3 cols corresponding to 3 main targets (e.g. exponential growth rate etc.). Actual values as scatter plot, including data forward of the determined inference time used. The posterior inferred values over times horizons are plotted as ribbon plot quantiles with colour determining the inference model.
• Fig 2: Overview: This fig aims at presenting the nowcasting (e.g. 0 horizon estimate) at rolling inference time points for each scenario with each inference model choice and possible misspecification of generation interval. Time horizon choice: Chosen horizon = 0 to align with Fig 1 but with other horizons as SI plots. Plotting details: 3 x 4 = 12 rows corresponding to 4 main scenarios (e.g. outbreak with measures etc.) and 3 main targets (e.g. exponential growth rate etc), the scenario GI is fixed to the middle mean GI (10 days; others are in SI) and 3 columns corresponding to underestimating mean GI (left), good estimation of GI (middle) and over estimating mean GI (right). Actual values as scatter plot. The posterior inferred value at the estimation date are plotted as boxplot plot quantiles¹ with colour determining the inference model.
• Fig 3: Overview: This figure concerns inference model scoring. Model score groups: scenario, inference model, horizon, inference date, and GI mis-specification (expressed as $\langle g \rangle - \langle G \rangle$ where $g$ is the used generation interval and $G$ is the scenario truth generation interval). Plotting details: TBD, first pass is the scoring data frame table. Second pass is some kind of stacked barplot.

Footnotes

1. @seabbs finds it more common to plot a rolling nowcast as a ribbon plot, without strong favour to that approach. I think this blurs the idea of how new information also changes past inferences. I'd prefer to have a ribbon plot as representing the inference over time made by a model at some inference time $T$, and have a sequence of box plots as representing what the right now estimate of the model at a sequence of $T$ values. The idea here is to keep these similar, but distinct, concepts separate visually. ↩

[seabbs]

Nice. I edited 1. to reflect my understanding of what we discussed (that we would plot all horizons (+ and -) for a single estimation date). Do you agree or did you have a different takeaway?

[SamuelBrand1]

No that makes sense, and was what I intended. I've done (bolded) edit for clarity.

[seabbs]

Do you think we have this plan out/moved to their own issues sufficiently to close this?

[SamuelBrand1]

Yes.