Homogeneity

Study guide: Exam 5

Homogeneity is the property of a group of risks whose expected loss costs are similar enough that their experience can be averaged into one rate — or projected with one development pattern — without material distortion.

$\text{Homogeneous if } E[\text{Loss Cost}_i] \approx E[\text{Loss Cost}_j] \text{ for all } i,j \text{ in the group}$

Homogeneity is always traded against Credibility. Splitting a book finer makes each group more alike but leaves each with less data; combining groups buys volume at the price of averaging together risks that genuinely differ. Every segmentation decision in ratemaking and reserving is a point on that curve.
In ratemaking, a non-homogeneous class means systematic cross-subsidy: the better risks in the class are overcharged and leave (see Classification Ratemaking), while the worse risks stay — the adverse selection spiral that risk classification exists to prevent.
In reserving, Friedland groups claims on the behaviour that drives development — reporting patterns, settlement patterns, how well case reserves can be set, the likelihood of reopening — alongside consistency of coverage, claim volume and severity. Two coverages may have similar loss ratios and completely different reporting and settlement speeds; combining them into one triangle produces development factors that describe neither. See Reserving Data Organization.
Werner's criteria for evaluating a rating variable (after Finger's Risk Classification) are statistical (statistical significance, homogeneity, credibility), operational (objective, inexpensive to administer, verifiable), social (affordability, causality, controllability, privacy concerns) and legal — homogeneity is one requirement inside the statistical criteria, not the whole test.
Homogeneity is checked, not assumed: compare loss ratios, frequencies and severities across subgroups within a proposed class, and look for a consistent difference that persists across years.

Testing a Class for Homogeneity A commercial auto class "light trucks" is rated at a single rate. Splitting it by radius of operation gives:

Subgroup Earned exposures Losses Pure premium

Local (under $50$ mi) $6{,}000$ $\$3{,}600{,}000$ $\$600$

Intermediate $2{,}500$ $\$2{,}125{,}000$ $\$850$

Long haul $1{,}500$ $\$2{,}250{,}000$ $\$1{,}500$

Is the class homogeneous, and what should be done?

Answer $\text{Class pure premium} = \frac{\$7{,}975{,}000}{10{,}000} = \$797.50$

Relativities to the class average: $0.75$ , $1.07$ and $1.88$ . Long-haul risks cost more than twice what local risks cost, so the class is not homogeneous — the single rate is $33\%$ above the local operators' cost and $47\%$ below the long-haul cost.

The remedy is to introduce radius of operation as a rating variable. The subgroup volumes support it: even the $1{,}500$ -exposure long-haul cell carries a difference far too large to be noise, and the relativities should be credibility-weighted against the class average rather than used raw.

If the insurer does not split the class, competitors that do will quote the local operators below its rate and take them, leaving it with the long-haul risks at an inadequate rate.

Subgroup	Earned exposures	Losses	Pure premium
Local (under $50$ mi)	$6{,}000$	$\$3{,}600{,}000$	$\$600$
Intermediate	$2{,}500$	$\$2{,}125{,}000$	$\$850$
Long haul	$1{,}500$	$\$2{,}250{,}000$	$\$1{,}500$

Homogeneity in a Reserving Segmentation An actuary must reserve a $\$30{,}000{,}000$ commercial package book. Property claims (60% of premium) close within a year; liability claims (40%) take six years. One combined triangle gives a $12$ – $24$ factor of $1.28$ .

Is the combined triangle acceptable?

Answer No — for two reasons, and the second is the one that bites.

First, $1.28$ describes neither coverage: property alone would be near $1.05$ , liability near $1.60$ . The blended factor is only correct while the mix stays exactly where it was.

Second, that is precisely what will not hold. If the insurer grows the liability side to $55\%$ of premium, the true blended factor rises, but the historical triangle still reports $1.28$ — the estimate is now biased low, and nothing in the triangle reveals why. This is the mix-of-business distortion.

Split into property and liability triangles. Each is homogeneous in emergence pattern, each has enough volume to be credible at this premium size, and the blended answer falls out of the sum rather than being baked into the factors.