# Actuarial Science and Risk Classification Insurance works by pooling. Individually unpredictable losses become collectively predictable — the law of large numbers, applied commercially. Actuarial science is the practice of estimating expected loss and setting premiums accordingly. The core technical activity is **classification**: sorting policyholders into groups whose expected losses differ, and pricing each group separately. ## The tension at the center of the field Classification and pooling work against each other, and the field has never resolved it. **Perfect classification destroys insurance.** If every individual's loss could be predicted exactly, everyone would be charged their own loss and no risk would be transferred. Insurance exists only because of residual uncertainty. Every improvement in predictive accuracy shrinks the pool that uncertainty creates. **No classification also destroys insurance**, through adverse selection. Charge everyone the average and low-risk buyers exit, raising the average, driving more exits — the classic death spiral. So the industry sits permanently between two failure modes, and where it sits is set by regulation, not by mathematics. **Where the line falls is a political decision that actuarial science can inform but cannot make**, which is a sentence the industry has spent a century avoiding saying plainly. ## What gets banned, and the pattern in what doesn't Some classifications are prohibited despite being predictive: - **Race** in essentially all lines, in the United States. - **Genetic information** in health insurance under GINA (2008) — enacted specifically because the classification would have worked. - **Pre-existing conditions** in ACA-regulated health insurance, a deliberate suppression of the most predictive available variable. - **Sex** in many contexts and jurisdictions, though treatment varies widely by line and by state. The pattern in the bans is not "these variables don't predict." They predict well. The pattern is a judgment that **some risks should be borne collectively rather than priced individually** — that the point of health insurance is precisely to protect people from the consequences of characteristics they did not choose. Meanwhile, credit-based insurance scores, ZIP code, occupation, education, and homeownership remain widely permitted, and every one of them correlates strongly with race. See [[Statistical Discrimination]] and [[Redlining and Housing Discrimination]] — the historical redlining maps were an actuarial product before they were a lending practice, and their geographic footprint persists in current risk classification through the ZIP-code channel. ## The proxy problem, stated precisely Prohibiting a variable does not remove its influence when correlated variables remain. A model denied race but given ZIP code, credit history, and vehicle type reconstructs much of the racial signal, and does so without anyone intending it. Modern machine-learning underwriting makes this sharper rather than milder. A model with hundreds of features finds proxy structure automatically — that is what fitting a model *is*. **The prohibition operates on the input list; the discrimination operates in the joint distribution.** This is the technical reason input-blindness is an inadequate fairness guarantee, and it connects directly to [[Fairness Impossibility Results]]. ## Where it is heading Two live pressures: **Genomics.** GINA covers health insurance and employment. Life, disability, and long-term-care insurance are substantially outside it. As predictive genomics improves, a large category of coverage becomes classifiable at the individual level — collapsing the pool for exactly the people most in need of it. **Climate.** Property insurers are repricing and withdrawing from wildfire and flood-exposed regions. This is actuarially correct: expected losses genuinely rose. It also means the insurance market is delivering an unambiguous verdict on [[Climate Change]] that the political system continues to contest — **the actuaries priced it while the legislature debated whether it was happening.** The resulting affordability crisis then arrives as a housing and municipal-finance problem rather than a climate one, which obscures the causal chain. ## Related [[Statistical Discrimination]] · [[Credit Scoring]] · [[Fairness Impossibility Results]] · [[The Value of a Statistical Life]] · [[Redlining and Housing Discrimination]] · [[Climate Change]] · [[Mathematics and Power Hub]]