ianmtaylor1 / efl Goto Github PK

Add predictplot() method(s) to plot, based on PPD samples,

• Game result distributions
• Game score distributions (for goal models)
• Game margin distributions (for goal models)
• Game total goals distributions (for goal models)

These should be identical to the functionality that currently exists by adding GameResult, GameScore, GameMargin, or GameTotalGoals stats to a Predictor. i.e. the plots should look the same, and the tails should be binned as appropriate.

Also consider:

• should there be a method for goal models that will put all four plots in a 2x2 grid for a particular game?
• should these be methods of models, or exterior functions that take dataframes output from the predict() method?

Let ArviZ do as much as possible, potentially posterior summaries and plots.

Option 1: model classes contain a to_arviz() method that returns InferenceData.
Option 2: model classes are themselves subclasses of InferenceData.

Probably needs to wait until ArviZ >= 0.6.2, since that version will add the predictions parameter in the from_pystan() function.

The aspect ratios of plots can be off. I noticed this when plotting game predictions (result, score, margin, total goals) in a 2x2 grid for multiple games.

• For now, restrict dependency to pystan version 2.x
• Keep an eye on how the pystan api will change in version 3.