What kind of Predictive Models do you build? Ground-up or Residual?

I am interested in finding out what the standard way of building predictive models is in the commercial line P&C industry (in North America).



I've been debating with colleagues about whether a ground-up model will outperform a residual model, or vice versa. By ground-up model I mean the model is applied to exposures directly, and a residual model is a model that's built and applied on top of the ISO/NCCI Manual Premium, so the model result is "deviations off of manual".



There are pros and cons to each method, so I am really curious to see what the norm is. Maybe I'll open up a poll when I figure out how that works. Appreciate your feedback! :rimshot:

What kind of Predictive Models do you build? Ground-up or Residual?