Grouping allows you to combine values or ranges based on the similarity of predictors' behavior.
There are no hard and fast rules for grouping. The profile should be smooth, the differences in behavior should be meaningful and the number of cases should be reliable. the Analytics Center portal uses sensible defaults, but some experimentation is usually worthwhile for producing more predictive power and reliability. The predicted behavior should not to be monotonic, too flat, or jagged without a good reason.
If the profile is too flat, try increasing the level of detail (granularity), setting a maximum probability of error and/or decreasing the minimum size (the level of evidence required for the behavior of a bin to be judged as representative). The number of bins should increase and the profile should become more varied.
If the profile is too jagged, try decreasing the level of detail and/or increasing the minimum size. The number of score bands should decrease and the profile should become smoother.