Experience rating

Why does an insurer whose whole business is pooling risk turn around and charge each customer by their own claims history?

An insurer exists to detach what you pay from what happens to you: a predictable premium in place of an unpredictable loss. Then it introduces a no-claims discount and the arrangement quietly reverses — have a bad year and you pay more for the next five. Your cost has been reattached to your outcome, which is the thing you bought cover to escape.

Is this the insurer taking back with one hand what it sold with the other? Or does the pooled contract destroy something only a personal price can rebuild?

Begin with what full cover does to behaviour. If a loss costs you nothing, care stops paying for itself. Not because people become reckless, but at the margin: the extra lock, the slower approach to the junction, the safety officer not hired. Each costs something and now saves nothing. This is hidden action, moral hazard, and it is distinct from the hidden-information problem worked through under adverse selection: there the question is who buys, before the contract; here it is what they do, after it.

How would you price care you cannot observe? You cannot watch a million drivers, but you can observe the consequence. If this year's claims move next year's premium, careless behaviour today has a price today — delayed, but real. Experience rating charges for an unobservable action by pricing its visible residue.

There is a second reason, and confusing the two is the commonest error. Suppose everyone's care is fixed and only their underlying riskiness differs, in ways no questionnaire captures. Claims history is then not a report on behaviour but evidence about type. Charging by it sorts the book, so low risks are not overcharged and do not leave — the defence against a pooled premium spiralling upward as the cheapest customers withdraw. One instrument, two quite different jobs: disciplining actions, and sorting people.

Can data tell them apart? Only with difficulty. A driver who claims less after being rated may have become careful, or may have been low-risk all along and is only now revealed. Separating incentive effects from selection effects in insurance data is a hard and still-argued problem in the field.

So how much weight should a personal history carry? The actuarial answer is precise. Your own record is informative but noisy; the class average is stable but not about you. So blend them by how much data you have. The standard limited-fluctuation form sets the weight Z on your own experience as Z = n / (n + k), where n is your years of exposure and k a constant chosen by judgement. Take k = 7 and a driver with three years of record: Z = 3 / (3 + 7) = 0.30. If that driver has averaged 0.20 claims a year in a class averaging 0.10, the rated frequency is 0.30 x 0.20 + 0.70 x 0.10 = 0.13, not 0.20. Thirty years of record gives Z = 30/37, about 0.81, and they would be charged close to their own number.

Notice what that formula concedes: most of a short history is noise, and charging you your own experience would be charging you for randomness. Pooling has not been abandoned but made partial, by an amount proportional to how much the insurer actually knows about you.

That gives the falsification test. If experience rating works through incentives, the sharpest prediction is not about average claim counts but about the margin: where the future premium penalty exceeds the small claim recovered, claims should stop being reported, and the effect should be strongest where the surcharge is steepest. If frequency were unchanged after a steep bonus-malus scale was introduced and only the composition of the book shifted, experience rating would be pure sorting with no incentive effect, and half the account above would be wrong.

Three refinements connect these pieces, and one of them is uncomfortable.

First, the asymmetry in bonus-malus scales is deliberate. A discount is earned one step per clean year and lost several per claim, so one claim is expensive for years — which is what makes the signal large enough to register against a small immediate loss.

Second, the uncomfortable part: the same design manufactures concealment. A customer whose claim would cost more in future premium than it recovers absorbs the loss and says nothing, so the insurer's own data understates true frequency and part of what the scheme buys is not more care but less reporting. In employer schemes the analogue is well documented as far as I recall: rated employers do see claim costs fall, but part of that comes from managing and contesting claims rather than from fewer injuries. Some of this is cost-shifting rather than safety, and the two are hard to tell apart from outside.

Third, who bears what. The insured bears the future price of today's misfortune — including misfortune that was nobody's fault, since a genuinely low-risk person still has claims occasionally and the scale cannot tell bad luck from bad driving. The insurer chooses the credibility constant, the class boundaries and the steepness of the malus, and the customer sees none of it. And where regulators forbid rating on some variable, the weight does not vanish; it migrates onto whatever remains observable, usually experience — so a rule meant to stop pricing by category ends up pricing more heavily by history, which correlates with the same categories anyway.

• How deductibles and experience rating overlap, and why an insurer usually wants both rather than one.

Every term the collection defines is gathered in the glossary.