No actuary has ever observed the worst earthquake or the most severe hurricane their portfolio could face. Unlike mortality or auto claim frequency, natural catastrophes are rare, heavy-tailed and, increasingly, nonstationary because of climate change. Modeling what has almost never happened, but could, is one of the most demanding challenges in actuarial work, and that is precisely where artificial intelligence is starting to push out the frontier of what can be calculated.
The heavy-tail problem
Ordinary claims experience can be described with abundant data and well-understood frequency and severity models. Catastrophes cannot. A single event can wipe out years of premium in a matter of hours, and the historical record rarely holds enough extreme cases to estimate the tail of the distribution directly. Extreme value theory provides the scaffolding: above a high threshold \(u\), excesses tend toward a generalized Pareto distribution,
$$P(X-u\le y\mid X>u)=1-\left(1+\frac{\xi\,y}{\sigma}\right)^{-1/\xi},$$
where the shape parameter \(\xi\) governs the heaviness of the tail. Estimating \(\xi\) from scarce data is delicate, and here artificial intelligence does not replace theory: it complements it, adding structure where the historical record is silent.
Catastrophe models: physics plus statistics
Since the 1990s the industry has assessed catastrophe risk with models that combine three modules. A stochastic event catalog generates thousands of plausible storms, earthquakes or floods; a hazard module translates each event into local intensity; and a vulnerability module turns that intensity into damage to the insured exposure. Vendors such as Verisk (formerly AIR) and Moody's RMS popularized this framework. The output is summarized in the loss exceedance curve, or EP curve, which assigns a return period to each loss level and from which metrics such as probable maximum loss and average annual loss are derived:
$$\text{AAL}=\int_{0}^{\infty}\ell\,f(\ell)\,d\ell.$$
Artificial intelligence fits into every module. It emulates expensive physical processes, refines vulnerability functions and enriches the description of exposure, three fronts where traditional error has been large.
Learning the climate, not just consulting it
Global climate models run at resolutions far too coarse to price a specific portfolio. Statistical downscaling with neural networks translates their output to the scale of a municipality or a postal code, capturing nonlinear relationships between atmospheric variables and local rainfall or wind. Beyond that, machine learning emulators, known as surrogate models, reproduce in seconds what takes a physical model hours. That makes it feasible to simulate thousands of climate scenarios and carry their uncertainty all the way through to the premium. Recent advances in AI weather forecasting, which can match or beat numerical models at certain horizons, hint at where hazard modeling is heading.
Seeing exposure from space
Much of the error in catastrophe modeling lies not in the hazard but in exposure and vulnerability, that is, in not knowing precisely what is insured, where, and with which construction characteristics. Computer vision applied to satellite and aerial imagery identifies roof type, number of stories, materials and proximity to bodies of water, sharpening the damage functions. After an event, the same techniques speed up loss estimation and the detection of affected areas, shortening the claims cycle exactly when speed matters.
The tail, reinsurance and capital
For the reinsurer and for the capital function, the tail is nearly everything. Catastrophe excess of loss covers, catastrophe bonds and capital structures under Solvency II or under the Mexican solvency capital requirement (RCS) all depend on carefully estimating return periods of one in two hundred years or more. A small bias in \(\xi\) or in the frequency of severe events is amplified in the tail and can leave an entity undercapitalized or overcapitalized. AI helps narrow that uncertainty, but it also introduces its own: a model trained on the climate of the past inherits a nonstationarity that often goes unacknowledged.
Governing what has no precedent
Climate change breaks a comfortable assumption: that the past is representative of the future. An AI model fitted to historical series can systematically underestimate a risk that is intensifying. That is why actuarial judgment is not dispensable; it is more necessary than ever. The Society of Actuaries maintains an active line of research on climate risk; the Casualty Actuarial Society works closely on catastrophe modeling and fairness in its use; and the Institute and Faculty of Actuaries has been emphatic in warning about climate tipping points and in calling for scenarios that do not ignore the worst outcomes. Validating an AI-supported catastrophe model calls for several disciplines:
- Separate the uncertainties: distinguish aleatory uncertainty, inherent in climate variability, from epistemic uncertainty, which reflects our ignorance of the correct model.
- Test the tail: subject results to stress tests and adverse scenarios, not just to the central fit.
- Document assumptions: keep an audit trail of data, assumptions and limitations for audit and regulatory review.
- Watch for drift: monitor that performance does not degrade as the climate shifts underneath the model.
The actuary in the face of the extreme
Artificial intelligence does not make the incalculable calculable, but it does move the frontier. It lets us see more plausible events, map exposure better and react faster when disaster strikes. What it cannot do is take responsibility for the number. Deciding how much prudence to add to a tail no one has observed, how much capital to hold for the event that has not yet happened, and how to communicate that uncertainty to a board or a regulator remains actuarial work. When facing extreme risk, the best model is not the most sophisticated one, but the one paired with judgment that knows when to distrust it.