The hidden assumption in plain Bühlmann
In the previous guide you met Bühlmann credibility and its tidy verdict: trust your own data with weight Z = n / (n + k), where n is how many observations you have and k is the ratio of within-risk noise to between-risk signal. There was a quiet assumption hiding in that little n, though. It counted observations as if each one were the same size — one driver, one year, one identical unit of risk. That is fine for a textbook fleet of identical drivers, but it is almost never true of real business.
Picture an insurer's book of group health plans. One employer covers 5,000 lives; the corner bakery next door covers 8. One year you wrote a plan for nine months, the next for a full twelve. A claims average built on 5,000 lives over a full year is a steady, trustworthy thing; the same statistic from 8 lives over nine months is a coin-flip. Plain Bühlmann has no way to say so — to it, both are simply 'one observation' and both would push Z the same way. We need a model that knows that bigger observations are quieter and ought to count for more.
Exposure: the right unit of size
The cure is to measure size honestly, and the actuary's word for size is exposure: the count of risk units actually at risk during a period — life-years, car-years, payroll dollars, number of policies. A plan covering 5,000 lives for a full year carries 5,000 life-years of exposure; the 8-life bakery for nine months carries 8 × 0.75 = 6 life-years. Exposure is the natural ruler because the random scatter of an average shrinks as exposure grows: more lives, less luck.
The Bühlmann-Straub model is simply Bühlmann rebuilt on exposure instead of a naive count. It does two things. First, when it summarises a risk's own history into a single number, it uses an exposure-weighted average — a year with 5,000 life-years pulls that average far harder than a year with 6. Second, it replaces n in the credibility weight with the total exposure m, giving Z = m / (m + k), keeping the very same k = EPV / VHM you already know. Everything you learned about k carries over untouched; only the meaning of the number on top has grown up.
Working a credibility premium
Let us price next year for one commercial group, the way an actuary actually does it. Suppose the empirical-Bayes step has already given us k = 800 (in life-years) and a collective mean of 1,000 per life — the rate for a brand-new account with no record of its own. Our group has accumulated 3,200 life-years of exposure, and over that history its own exposure-weighted average claim has run at 1,150 per life. The question is the same as ever: how much of that 1,150 do we trust?
Z = m / (m + k) = 3200 / (3200 + 800) = 0.80
credibility premium = Z * own + (1 - Z) * collective
= 0.80 * 1150 + 0.20 * 1000 = 1120 per lifeThat 1,120 is the credibility premium — the exposure-weighted blend of the group's own 1,150 and the collective 1,000. Now run the same arithmetic for the tiny bakery, with only 200 life-years: its Z is 200 / (200 + 800) = 0.20, so 80 percent of its rate comes from the collective and only a fifth from its own thin, noisy record. Exposure has done exactly what we wanted — the large, information-rich account speaks mostly for itself, while the small one is gently but firmly pulled toward the crowd.
Where it earns its keep: experience rating
Bühlmann-Straub is the workhorse behind experience rating of commercial accounts. When a group's policy comes up for renewal, the underwriter does not invent a number; a credibility engine blends the account's own loss history with the manual rate for its class, and the weight is set by how much exposure the account brings. A 5,000-life employer with three clean years sees that good record reflected in a discount it genuinely earned. A small shop with one freak claim is not punished for a single unlucky year, because its low Z keeps it tethered to the class rate.
The same machine shows up wherever exposures vary widely from risk to risk and year to year: workers' compensation experience modifications, group life and health renewals, and reinsurance, where a treaty's exposure can swing enormously between accounts. In each case the parameters k and the collective mean are estimated once across the whole portfolio — the empirical-Bayes step you met earlier — and then every individual risk is rated by its own exposure. One structural model, many tailored premiums.
- Estimate the structural parameters (EPV, VHM, hence k) and the collective mean once, across the whole portfolio.
- For each account, total its exposure m and form its exposure-weighted average claim.
- Set Z = m / (m + k); a bigger account automatically earns a higher Z.
- Blend: credibility premium = Z × own average + (1 − Z) × collective mean.
Honest limits
Bühlmann-Straub leans on one structural assumption worth stating plainly: it assumes each risk's process variance is inversely proportional to its exposure, the clean Poisson-style scaling where doubling the lives halves the variance of the average. That holds beautifully for independent, homogeneous risk units. It frays when the units are not independent — a single factory fire can hit hundreds of 'lives' at once, so a 5,000-life account is not really 5,000 independent coin flips, and its true volatility shrinks slower than the model assumes. When that happens, the model hands a large account more credibility than it has truly earned.
Two more honest cautions. First, the exposure measure must be the right one — payroll, headcount, sales — and measured correctly; if an account is 'large' only on paper, its high Z will confidently mislead. Second, what comes out is still a credibility premium: a best estimate of expected loss, not a finished price. It must still be loaded for expenses, profit and risk margin before it becomes a gross premium anyone can sell. The credibility blend answers 'what will this cost?', not 'what should we charge?' — those are different questions.