Part 8: Understanding normalise_to

Part 8: Understanding normalise_to

The normalise_to parameter controls how the relativities are scaled. In a new cell, type this and run it (Shift+Enter):

rels_mean = sr.extract_relativities(normalise_to="mean")

print("Area relativities, normalised to exposure-weighted mean:")
area_mean = rels_mean[rels_mean["feature"] == "area"].sort_values("level")
print(area_mean[["level", "relativity"]].to_string(index=False))

print("\nArea relativities, normalised to base level A:")
area_base = rels[rels["feature"] == "area"].sort_values("level")
print(area_base[["level", "relativity"]].to_string(index=False))

You will see two different-looking tables for area, even though the underlying model has not changed.

normalise_to="base_level" sets the named base level to exactly 1.000. All other levels are expressed relative to it. This matches GLM convention and is what you use for Radar imports and committee presentations.

normalise_to="mean" sets the exposure-weighted portfolio mean to 1.000. Levels above 1.0 are above-average risk; levels below 1.0 are below-average. This is useful for internal portfolio analysis but is harder to compare to a GLM.

Rule of thumb: use base_level when comparing to a GLM or importing to a rating system. Use mean when presenting to an underwriting team who thinks in terms of "above average" and "below average" risk.