Hi Ed 
nice to see you, and for the question.
Trying to predict churn usually only works well when it’s done in a tailored way for a specific company.
One technique is to use segmentation to identify customer types who are more prone to churn due to weaker PMF for that segment (or some other characteristic of that segment). For example we know our PMF is weaker for subscription box companies than it is for SaaS companies, so we know going into those relationships that there’s a higher risk we lose them, even if we don’t see any specific user behavior that would indicate they are thinking about cancelling.
The biggest red flags of course are when people stop using your product in some important way, e.g. for us it might be that they stop logging in, or remove their data from their account - but at this point things are usually too late to turn around. Predicting churn before you see one of those serious alarms is very tough. The best technique we’ve found is to do quarterly account check-ins, as these calls often surface issues that haven’t yet bubbled up into any sort of action.
One thing we notice is that people tend to go quiet before they churn, it’s often not the ones who are screaming at you and threatening to churn, they just want you to solve whatever issue they have. The ones who have already decided to leave don’t bother talking to you.
It’s also the case that the customers that don’t respond to your NPS surveys have a higher churn rate than the ones who give you a low score on the NPS survey.