In a previous post, I talked about credibility problems arising from having too few losses in a policy year. A related but distinct problem arises when a few large losses dominate a dataset.
In many respects, the easiest data to work with is a large group of losses of similar size. A large private passenger automobile book is a great example. There are bigger claims and smaller claims, but the bulk are small, and are in a pretty predictable range.
The polar opposite might be a small book of medical malpractice claims. Though such claims can be small, they can also easily range into seven and eight figures. Worker’s comp, especially for higher-risk businesses, can be similar.
Multi-million dollar losses would seem like enough to worry about. But no; there’s another problem here. When one or two large claims occur in a year with just a few smaller claims, the large claims are pushing the numbers around by themselves. Now it can become difficult to tell which claims are typical: the large ones or the small ones? And small claims, which might be settled privately and quickly, play out very differently than large claims, which have to wind slowly through the courts.
For those of us charged with estimating the likely future value of these losses, there are a few approaches available. One is to pull the large losses out of the data altogether. This makes sense when the large losses can be argued to be truly different from the small ones, and when there is good reason to think they will not be repeated. (Though the business side and the audit side often have different opinions about this!)
Another is to choose appropriate caps on the data. If development of losses on large claims is capped at a low level, their influence on the data can be diminished, and more fundamental patterns show up clearly.
Large losses may not jump out in the same way as zero-dollar years or single-digit claim counts. But they can have a huge impact on the bottom line, so it’s essential to keep them in mind.