Yes, generally speaking, there is a correlation between usage and likelihood to stay on and not cancel. When we do our analysis, we're starting at the unique event level. I can't think of the data science term off the top of my head for it but our process doesn't require events to be keyed in, so while we have an event called 'login' our platform thinks of it as 'event 1' --- there is no semantic meaning to the event to our system, and this is intentional to allow us to test all events without knowing exactly what they are and under what circumstances they fire off.
>or is it that customers which have multiple users using the SaaS application are more likely to stay on ?
This is an interesting question, and I wouldn't say this is true. There is a lot of complexity when an account has 10 users, and 2 are super power users, active several times a day, when the other 8 are different degrees of engaged. Maybe a few never even registered in the first place. We have several methodologies for 'rolling up' user associated with one account to the account level which is key for our customers where we're not scoring likelihood of users to cancel, but likelihood of accounts to cancel.
> What signals are emitted by the customers who are about to discontinue ?
Varies for all of our customers. One thing to consider is we make sure we're not contaminating our dataset... if you have an event which is called "customer clicked don't renew button" that is helpful to you but we want to predict weeks or months before they click that button, and not take those types of events into account for training purposes.