The actuarial table

For most of history, the price of a life was an argument. If you wished to sell someone an annuity — a stream of income for as long as they lived — you had to guess how long that would be, and the guess was a matter of appraisal, temperament, and nerve. A seller who guessed too generously was ruined when the buyer outlived him; a seller who guessed too meanly cheated the prudent and drove them away. Governments sold life annuities for centuries with almost no regard to the age of the buyer, pricing a life of twenty and a life of seventy at nearly the same rate, because there was no disciplined way to say how the two differed. The knowledge did not exist. What existed was opinion, and opinion, dressed in confidence, was expensive.

The thing that replaced the opinion did not arrive as a prophecy. It arrived as bookkeeping. In 1662 a London tradesman named John Graunt published his Natural and Political Observations upon the Bills of Mortality, a slim study of the weekly death registers that the city had kept, unloved, for decades. Nobody had thought to read them as data. Graunt did. By tallying the causes and counts across years, he could show regularities no single parish clerk could see — that the ratio of male to female births held steady, that certain diseases carried off a stable fraction each year, that the record of deaths, aggregated, had a shape. He even attempted a crude table of how many of a hundred people born survived to each age. It was rough, and he knew it was rough. What was new was the method: a claim about human life derived from the recorded fact of who had died, and offered as something a reader could check against the registers themselves.

Three decades later, in 1693, Edmond Halley took the next step with better raw material. Working from unusually complete birth and death records for the city of Breslau, he built a life table — a column of figures giving, for each age, the number still living out of an original cohort — and then did the thing the merchants could never do honestly: he used it to price an annuity. Given the table, the fair value of a life income at any age became a calculation rather than a wager, because the table said what fraction of buyers, on the evidence, would still be drawing the income in each future year. Halley was explicit that his table did not tell you when any particular person would die. It told you how a great many people, taken together, had died, and that aggregate was enough to price the risk without pretending to see the individual future.

The table never claimed to know when you would die. It claimed, and could prove, to know how often people like you had — and that honesty was the whole of its authority.

Honesty about uncertainty

This is the achievement worth dwelling on, because it is easy to mistake it for a lesser one. The mortality table did not make death predictable. A person is not an average; the individual case remained, and remains, opaque. What the table did was replace a confident opinion about the individual with a disciplined estimate about the aggregate — and, crucially, it showed its work. The estimate was built from a stated record, by a stated method, so that a skeptic could ask where the numbers came from and be answered. Graunt's registers were public. Halley's Breslau data and his arithmetic were laid out for anyone to redo. The trust the table earned was not the trust of a seer who turns out to be right; it was the humbler, sturdier trust of a clerk who shows you the ledger and invites you to add it up yourself.

Over the following century and a half this discipline hardened into a profession. Life assurance societies formed, priced their policies against mortality tables refined from their own accumulating experience, and learned — sometimes through insolvency — that a table built on wishful data or applied beyond the population it was drawn from would eventually collapse. The actuary emerged as a particular kind of expert: not one who claims to know the future, but one whose whole art is quantifying how much is not known, and pricing accordingly. The competence and the humility were the same act. To price a risk honestly is to state, precisely, the boundary of your own knowledge.

Calibrated humility

What the actuary produces, at its best, is a probability that knows it is a probability. A well-made table does not say this man will live to seventy; it says that of men entering at his age, the recorded evidence puts a certain fraction at the far side of seventy, and it can tell you how confident that figure is and on what body of experience it rests. This is calibration in the exact sense: when the table says a fifth of a cohort will not survive the decade, about a fifth should not, and if reality diverges, the divergence is a finding that sends the actuary back to the data. The number is answerable to the world. It is checked against outcomes, revised when it drifts, and carried always with a statement of the population it was drawn from and therefore the population to which it may honestly be applied.

Notice how much of this is transparency rather than cleverness. The source of the actuary's authority is not a superior faculty of foresight; it is the willingness to expose the record and the method so that the estimate can be criticized and improved. A table whose underlying data were secret, or whose calibration were merely asserted, would be worth exactly as much as the annuity-sellers' hunches it replaced — which is to say, nothing you could rely on. The virtue is a compound one: honesty about what the record contains, discipline in reasoning from it, and the restraint to treat a probability as a probability and not smuggle it across the line into a claim of certainty about the case in front of you.

An actuarial discipline for the decision

Machine decision systems are now pressed into precisely the actuary's job. A model that scores a loan applicant, flags a transaction, or ranks a résumé is doing what Halley did: pricing uncertainty about an individual from the recorded experience of many. This is not a criticism; it is the correct description of the task, and it is often a task worth doing. The trouble is that the actuary's discipline is frequently stripped out on the way. A score arrives presented as if it knew the individual, when at best it knows a rate. Its calibration is unstated — nobody has told you whether, when the system says seventy percent, the thing happens seventy percent of the time. The record it was built from is unshown, so you cannot ask whether the population you belong to is one the table was ever drawn from. The humility that made the mortality table trustworthy has been quietly discarded, and what remains is the confidence — the very thing the table was invented to replace.

The remedy is not to forbid the estimate but to restore the discipline around it, and the mortality table tells us exactly what that discipline owes. It owes a stated, checked calibration: a claim that the probabilities mean what they say, tested against outcomes and revised when they drift, not merely announced. It owes the provenance of its judgment — the record it was built from, and therefore the honest boundary of where it applies and where it does not. And it owes the restraint to treat a probability as a probability: a score that estimates a likelihood is not a verdict about a person, and a decision that lets it harden into one has committed the annuity-seller's original sin of pricing a life by conviction. A Decision Receipt that carries the estimate together with its calibration and its provenance turns a score back into what Halley made it — a defensible, contestable price on uncertainty, rather than an oracle's pronouncement.

Elsewhere in this series I have argued for the control group, the discipline that tests whether a plausible causal story is true. This is a different owed thing. The actuary is not testing a causal claim; the actuary is pricing an aggregate uncertainty honestly and showing the basis, so that the price can be checked and the humility preserved. Both are ways of refusing to let confidence stand in for evidence, but the actuarial refusal is the one a probabilistic machine most needs, because a machine that outputs a number is always tempted to let the number pretend it knows more than it does. Three and a half centuries ago a tradesman reading death registers and an astronomer pricing annuities worked out that the honest answer to an unknowable individual future is a calibrated statement about the many, offered with its evidence attached. We have built systems that make such statements at enormous scale and stripped the attachments off. The table is still the standard. The task is to make the machine as honest as the clerk.

— Dispatches · Summit Cognitive

---