I agreed, also I agree that the pro and con of each one isn't stated enough either. I recently attended Dr. Frank Harrell's workshop and he really put the diff

I agreed, also I agree that the pro and con of each one isn't stated enough either. I recently attended Dr. Frank Harrell's workshop and he really put the differences between the two from his experiences in perspective. He advocate more for continuous variable responses than what most ML does. He also put into context what ML does well where the "signal to noise" ratio is so that the noise are little and signal are more. But when the noises are getting bigger statistical model will do better. His examples is the titanic dataset. Paraphrasing but he stated, "I don't care who live or die." The better question is what is their tendency to live or die? Instead of classifying it into two categories why not give a percentage? What if a person is 49% likely to live? Aren't you playing God? And it is up to the person that's using the model to decide not the modeler. Another case is that it's a loss of information by forcing it into a classification problem. He went on how ML tend to do this force classification. And how it is very good at things that have low signal to noise, such as game of Go, speech rec, visual rec, etc... It's an interesting contrast than Dr. Leo Breiman's The Two Cultures paper. At least that's my take away, Dr. Breiman may be less opinionated? Both are very critical of statistic/data models but Dr. Frank Harrell have good remedies and counter points.