I want to create a binary model which predicts whether someone has improved his state. I am testing possible variables as explanatory variables in order to make some recommendations. Now my binary model is quite weak, AUC is around 0.6 and there are hardly any cases predicted above 0.5 in the test set.
Now I could simply improve my model by including the previous state which is highly correlated with improvement, but for me this is too obvious. It might 'kick out' other explanatory (more actionable) variables.
So now I am faced with this question of whether to go for accuracy or simpler/actionable model.
What do you recommend?