When the Average Doesn't Exist: Why Learned Insurance Payouts Fail Under Distribution Shift

Parametric insurance is the leading hope for extending climate cover to people conventional insurance cannot reach: a satellite reading crosses a threshold and the money moves, with no inspector and no delay. Everything then rests on how closely the index tracks the loss, and machine learning is being recruited to close that gap. We show that at the granularity where 44% of real contracts settle, the quantity being predicted has no average to predict. On the complete US federal flood record (2.1 million claims) an individual claim is heavy-tailed but tractable (Hill tail index α = 2.43). Aggregate to the county-month totals that area-index and sovereign products actually pay on, and α is 0.75. Below 1 there is no finite mean to estimate: the running average never settles over any range a claims history can observe, a single event holds 10% of all losses, and further pooling, across counties or states, does not recover it. Because this is a property of the target, it binds every model. Payment schemes defined through an expected loss, including the expectile-optimal designs proposed for this exact problem, are undefined here. For the quantile schemes that survive, we derive and verify a regret law capping how far a finite history lets a payout lean toward the policyholder before estimation error dominates. Can a model succeed regardless? We answer by elimination: six configurations, three model families, ten seeds, three regimes. In-distribution, gradient boosting genuinely works: it beats holding cash in 10 of 10 seeds and cuts basis risk by 19%. Under temporal or climate shift, nothing does, in any of 110 runs: not a budget constraint, a frequency–severity decomposition, quadrupled capacity, a seed ensemble, an oracle rescaling, or a real natural-capital dependency graph, which performs no better than a scrambled one. What makes the result actionable is its cause. The driver is not the hazard but the unevenness of the exposure being pooled: normalising by policies in force liftsˆ αfrom 0.71 to 1.00 on the matched subset. The binding constraint is therefore a contract parameter, not a fact about weather, i.e., the same losses, measured per unit of exposure, are estimable. The lever is therefore, we conclude, not the model; It is what a contract insures, and at what level it pools.