Description
Model the size and frequency of rare extreme values to estimate tail percentiles and return levels beyond the range of most data.
Averages and even 95th percentiles say little about the worst experiences. Extreme value analysis models the tail directly. The block-maxima approach takes the largest value in each block (day, week) and fits the generalised extreme value distribution. The peaks-over-threshold approach fits a generalised Pareto distribution to excesses over a high threshold, using more data and allowing covariates. Tail heaviness is summarised by the tail index, estimated by the Hill estimator from the k largest values and read from a plot against k in a stable region, with Pickands and moment estimators as alternatives. Diagnostics precede fitting: exponential and Pareto QQ plots and the mean excess plot (rising for heavy tails) show tail type and guide threshold choice. Extreme quantiles and return levels follow, with delta-method intervals. Dependent series need declustering via the extremal index.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Post-launch, Evaluation
- Data
- Quantitative
- Effort
- Days
Steps
- Check tail type with an exponential QQ plot, a Pareto QQ plot (log data) and a mean excess plot; a rising mean excess indicates a heavy tail.
- Either split the series into blocks (for example days or weeks) and fit the generalised extreme value distribution to block maxima by maximum likelihood, or choose a high threshold and fit the generalised Pareto distribution to excesses.
- Choose the threshold or number of top values k by plotting estimates of the tail index against k (Hill plot) and picking a region where they are stable.
- If exceedances cluster in time, estimate the extremal index and decluster before fitting.
- Check fit with QQ plots of the fitted tail and, for peaks over threshold, check that gaps between exceedances look exponential.
- Estimate extreme quantiles or return levels (for example the latency exceeded once a month) with intervals, and report sensitivity to block length or threshold.
Sources
- Analysis of Financial Time Series (Tsay, 2002)
- Statistics of Extremes: Theory and Applications (Beirlant et al., 2004)
- Order Statistics (David & Nagaraja, 2003)
Also known as
- Extreme value analysis
- EVT
- Extreme value theory
- Generalised extreme value fitting
- GEV
- Peaks over threshold
- POT
- Generalised Pareto fitting
- Hill estimator
- Hill plot
- Mean excess plot
- Return level estimation
Categories
- Analytics & Behaviour
- Statistics & Modelling