In context: this report covers a source published on 4 Nov 2026.

Group life insurers often price an employer's cover partly on its own claims history, sorting insured groups into risk groups that each carry a tariff factor. A preprint posted on arXiv on 4 November 2026 by Thom van Rijn, Frank Weber and Michael Merz, of the University of Hamburg Business School and the University of Basel, asks whether that sorting, often done by hand with arbitrary thresholds, can be automated. It has not been peer reviewed.

The test data are mortality records for 2014 to 2022 from a large European life insurer that the authors do not name. Experience is grouped into 84 economic sectors, with between about 36,700 and 42,600 contracts and 327,000 to 397,000 insured people each year. Many sectors are thin: the median sector has 1,743 insured and 1.79 expected claims a year. The authors compare two measures with expected values: how many claims occur and the size of the survivors' pensions paid.

Their method first applies a multidimensional Bühlmann-Straub credibility model, a standard actuarial tool that gives less weight to sparse data, to both measures. It then sorts the sectors into five risk groups with k-means clustering, a common machine-learning technique. Against k-means on raw experience, with or without outliers, over tariff years 2019 to 2025, a sector's tariff factor changed on average by 7.21% from one year to the next, compared with 16.60% and 14.46%. Changes of 10% or more were 23.57% of the total, against 60.71% and 52.62%. Group factors also varied less across 1,000 bootstrap resamples, and the method kept its overall lead with three or seven groups, though less uniformly with seven.

The paper does not test whether the groups predict later claims better; it judges them on differentiation, central alignment and stability. The authors write that tariff changes still 'remain substantial', mainly because sectors near group boundaries switch groups, and in some groups plain k-means kept more sectors in place. The results come from one insurer, the authors call the rule that merges the two measures 'a pragmatic choice', and the paper does not say whether the insurer uses the method. They suggest soft clustering as a next step.

Read the source

arXiv · Industry body · 4 Nov 2026. This is an original brief based on the linked publication. Company statements and forecasts are attributed to their source.

arXiv ↗