What a Performance Guarantee Costs, and How Long It Should Run

How to read this sheet

The reserve is an expected value and a guarantee is not settled in expectation, so treat this as the price of the promise rather than as next year’s claims bill. Real claims arrive lumpy, they cluster in the batches that were wrong, and a single bad lot will exceed the reserve on a whole year of good ones. That is an argument for holding the reserve rather than against computing it. Read the two inverted figures as the real output. The reserve depends entirely on a distribution that almost nobody has measured on their own goods, so the useful question is not what the reserve is but how wrong the estimate can be before the answer changes. A guarantee that needs a failure shape of three point two when the wash trial says two point four is not marginal, it is wrong, and the gap is the margin of safety in the laboratory work rather than in the price. Shape matters more than most people expect. At a shape of one, failures arrive at a constant rate and a term twice as long costs very nearly twice as much. At two and above they cluster, and the same doubling costs three or four times as much - which is why matching a competitor’s longer guarantee is rarely the small concession it appears to be, and why the concession should be priced before it is made rather than after. What is not modelled. Claims are not discounted, which overstates a long guarantee - the safe direction for a seller, but worth remembering on a twenty-five year design life where it is a large effect. Nothing here is a reliability growth model, so a product that is genuinely improving will be over-reserved on old data. Batch-to-batch variation is absent, and it is the largest real risk. Pro-rata guarantees, repair rather than replacement, and caps on liability all reduce the exposure materially and none of them are here; a capped guarantee is a different and usually much better instrument than the uncapped one this sheet prices.