I used to do this kind of work in the banking world, we built massive models to predict fraud in retail banking. The key observation - improvement in model performance was directly related to its invasiveness - i.e. the more you know about the client the more accurate and powerful the model became. In the end you end up tying together every piece of knowledge you have all the client or prospective client, across all databases and business channels and LOBs and building massive decision trees. But even at that level the false positive rate is very high - simply because fraud is a rare event as a proportion of the whole pool.

In the end I came to the conclusion that the blind pursuit of model performance and thereby business metrics in this realm was simply incompatible with the idea of privacy. One must be willing to philosophically accept a level of fraud just as a free society accepts a level of dissent or a parent a level of rebellion in her/his children.