Bayesian Logistic Regression

If you've ever searched for evaluation metrics to assess model accuracy, chances are that you found many different options to choose from. Accuracy is in some sense the holy grail of prediction so it's not at all surprising that the machine learning community spends a lot time thinking about it. In a world where more and more high-stake decisions are being automated, model accuracy is in fact a very valid concern. But does this recipe for model evaluation seem like a sound and complete approach to automated decision-making? Some would argue that we need to pay more attention to model uncertainty. No matter how many times you have cross-validated your model, the loss metric that it is being optimized against as well as its parameters and predictions remain inherently random variables.