Actuarial and Quantitative Risk
The Global Risks Alliance (GRA)
Extreme value theory gives a statistical basis for estimating the size and frequency of losses and hazards beyond what the record shows. It is used for large-loss pricing, catastrophe hazard, operational risk and capital at high confidence levels. Results depend on thresholds, sample size and whether conditions are stable. This page sets out how extreme value methods work, where they fail, the data they need and who relies on them.
Fig.Reserving workbench. Design wireframe with placeholder names and illustrative values, not a record or a live system.
Extreme value methods are statistical techniques for estimating the size and frequency of the largest losses and hazards, beyond the observed record. They rest on limiting distributions, threshold choices and stable conditions. Actuaries, capital model owners, validators and supervisors decide how they are used.
Which places drive loss in a severe year, and how far it reaches into reinsurance. Illustrative design values.
Tail estimation for hazards, markets and operational losses. The dossier below sets out how the risk develops, what is measured, where the data come from, how it is modelled and who decides.
Under broad conditions, block maxima follow a generalised extreme value distribution and exceedances over a high threshold follow a generalised Pareto distribution.
A high threshold respects the theory but leaves few points; a low threshold adds data but bias. Results shift with the choice.
Estimating a rare return period from a short record gives wide confidence intervals that are often left unreported.
If climate, exposure or claims practice change, past extremes may not represent future ones, and stationary fits mislead.
Each term as this page uses it. Where a standard-setting body defines the term, the wording follows that body.
| Term | Definition |
|---|---|
| Block maxima | The largest value in each fixed period, such as a year, used to fit a generalised extreme value distribution. |
| Generalised Pareto distribution | The distribution used to model exceedances over a high threshold. |
| Shape parameter | The parameter that determines whether a distribution's tail is heavy, light or bounded. |
Each indicator carries its unit, its cadence, the kind of source behind it and the evidence label a value would carry. No values are shown: a programme supplies them under its own authority.
| Indicator | Unit | Cadence | Source type | Label |
|---|---|---|---|---|
| Shape parameter of the tail | dimensionless | per analysis | Fitted extreme value models | MODELLED |
| Return level at stated return period | currency or hazard unit | per analysis | Fitted models | MODELLED |
| Confidence interval width for return level | currency or hazard unit | per analysis | Bootstrap or profile likelihood | MODELLED |
| Exceedances above threshold | count | per analysis | Data | OBSERVED |
| Record length | years | per dataset | Data | OBSERVED |
| Trend in annual maxima | change per decade | per analysis | Non-stationary fits | ESTIMATED |
The topic as a computational graph: named sources feed the methods, an evidence review gates the outputs, and the decisions stay with the authorities that hold them.
Fig. 02Nodes carry their type, two metric rows and a state; connectors show what feeds what; dashed frames name the owners; the timeline is the cadence. Design of the pipeline, not a record of a run.
Named public datasets and institutional records, each with its licence class. Sources by agreement are used only under their terms and never republished.
| Source | What it gives | Licence class |
|---|---|---|
| Coles, An Introduction to Statistical Modeling of Extreme Values | Standard reference text on extreme value methods | Proprietary, published book |
| R extRemes and evd packages | Open-source extreme value tools | Open source, GPL |
| NOAA and national meteorological hazard records | Long records of hazard extremes | Public domain or terms per service |
| Operational risk loss data consortia publications | Large operational loss records | Members only; summaries free to read |
| Large-loss databases of insurers and reinsurers | Individual large claims | Proprietary, internal |
Actuarial and Quantitative Risk prepares evidence. The decisions below are taken by the authorities named, under their own rules, and never by the platform.
| Who decides | The question | Evidence this page serves |
|---|---|---|
| Actuaries and risk modellers | Which extreme value method and threshold to use | Stability diagnostics and data review |
| Capital model owners | How tail estimates enter capital | Uncertainty and sensitivity evidence |
| Validators | Whether tail fits are adequate | Back-testing and alternative fits |
| Supervisors | Whether tail calibration supports regulatory capital | Model documentation and validation |
Two classical results underpin the field. The distribution of block maxima, such as the largest annual flood, tends towards the generalized extreme value distribution under broad conditions. The distribution of exceedances over a high threshold tends towards the generalized Pareto distribution. A single shape parameter determines whether the tail is heavy, light or bounded, and it is the most important and most uncertain quantity in any extreme value analysis.
Practical choices drive results: the block size or threshold, how to handle clusters of extremes that occur together, such as aftershocks or volatile trading days, and whether to include covariates for trends or seasonality. Non-stationary extreme value models let parameters change with time or with climate indices, which is essential when the underlying process is changing.
Multivariate and spatial extremes address joint behaviour, such as storm surge with rainfall, or losses across regions. Dependence in the extremes can differ from dependence in the bulk of the data, which has direct consequences for diversification.
Fitting the generalized Pareto distribution to exceedances; results depend on threshold choice.
Fitting the generalized extreme value distribution to block maxima; uses data less efficiently.
Letting parameters vary with time or climate; risks over-fitting short records.
Limits: estimates far beyond the data rest on the assumption that the limiting distribution applies, and confidence intervals for long return periods are wide. Reporting return levels without intervals overstates certainty.
The current service level is R0: information and enquiries. An enquiry receives a reference and names the next step; applications, payments, compute allocations and registrations open only when their team, terms and review capacity exist.