Artemis draw predictions

[calumrussell]

Artemis maps from Bivariate Poisson directly to outcome probabilities. The problem with this approach seems to an underprediction of draws.

First step is to quantify whether this is occurring due to an unprediction of zero outcomes (i.e. the standard Poisson problem which can be solved with zero inflation) or whether it is across the board underprediction of draws (i.e. can be solved with Dixon-Coles).

The possible solutions are (depending on first step):

1. Zero inflation
2. Dixon-Coles
3. Using off/def strength parameters as input to multinomial logistic regression (this is what Ares uses, and is basically a shortcut if everything else is going to take too much time).