The normal distribution describes the middle of the data well and the extremes badly. But risk lives in the extremes: the worst day of the year, the 1-in-200 storm. Extreme value theory is the probability of those worst cases, and the surprise is how little you need to assume about the rest of the distribution.
Simulate 1,000 years of 250 trading days, with daily losses of the same 1% volatility, and keep each year's single worst day: its block maximumBlock maximum: the largest value in a block of observations, such as the worst daily loss in a year or the largest claim in a year.. Do it twice: once with normal losses, once with fat-tailed Student t losses.
Normal daily losses Student t daily losses (ν = 4)
Yearly maximum daily loss, in %. 18 Student t years and 0 normal years lie beyond 9.5% and are off the chart.
The normal maxima huddle together: median 2.77%, and even the worst of 1,000 years only 4.75%. The Student t maxima have a median of 3.87% and a long right tail; the worst year reached 20.4%. Both histograms have the same family of shapes, the GEV distributionGeneralised extreme value (GEV) distribution: the only possible limit of suitably shifted and scaled block maxima. Parameters: location μ, scale σ, shape ξ., but with different values of one number, the shape ξShape parameter ξ: sets the tail. ξ > 0 heavy (power) tail, ξ = 0 light (exponential) tail, ξ < 0 a finite upper end..
The rules
Maxima. The distribution of the largest of n draws, and why it needs rescaling.
The GEV and its three cases.GumbelGumbel distribution: the GEV with ξ = 0. Maxima of light-tailed variables such as the normal end up here., FréchetFréchet case: the GEV with ξ > 0. Maxima of heavy-tailed variables with power tails, such as the Student t, end up here. and WeibullWeibull case (reversed Weibull): the GEV with ξ < 0, which has a finite upper end point μ − σ/ξ., with the tail indexTail index α = 1/ξ: for a power tail, P(X > x) falls roughly like x^(−α). Smaller α means a heavier tail; moments of order α and above are infinite. 1/ξ.
Return levels. The 1-in-T levelReturn level z_T: the level a yearly maximum exceeds with probability 1/T, the 'T-year' or '1-in-T' level. of a yearly maximum.
Exceedances over a threshold. The GPDGeneralised Pareto distribution (GPD): the limit distribution of excesses over a high threshold, P(Y > y) = (1 + ξy/β)^(−1/ξ). above a thresholdThreshold u: the level above which observations are treated as tail data in the peaks-over-threshold method. u.
The formal version
P(Mn≤x)=F(x)n;if P(anMn−bn≤x)→H(x) non-degenerate, then H=Hξ;μ,σHξ;μ,σ(x)=exp{−[1+ξσx−μ]+−1/ξ},ξ=0;H0;μ,σ(x)=exp{−e−(x−μ)/σ}zT=μ−ξσ[1−(−ln(1−T1))−ξ];P(X−u>y∣X>u)≈(1+βξy)−1/ξ,y≥0VaRq≈u+ξβ[(Nun(1−q))−ξ−1];1−F(x)=x−1/ξL(x),L slowly varying⇔Freˊchet domain,ξ>0
X₁, …, Xn are independent with distribution F and Mn is their maximum; an > 0 and bn are normalising constants; [·]₊ means the formula holds where the bracket is positive; u is a high threshold, β > 0 the GPD scale and Nu the number of the n observations above u; L is a slowly varying function (L(tx)/L(x) → 1 for every t > 0). Sources: Fisher and Tippett (1928); Coles (2001), chapters 3 and 4; Embrechts, Klüppelberg and Mikosch (1997); Pickands (1975); Balkema and de Haan (1974).
Try it
A GEV with location 0 and scale 1. Slide the shape from negative (a hard ceiling) through 0 (Gumbel) to positive (a heavy right tail), and watch the 1-in-T return level. At ξ = 0.25, the value that matches daily returns with Student t tails, the 1-in-200 level is far above the Gumbel's.
Try itShape the extreme value distribution
Return period T (years)
GEV density with μ = 0, σ = 1 (solid), the Gumbel ξ = 0 (dashed), the 1-in-100 return level (purple line) and the finite end of the support (red dot).
FamilyFréchet (ξ > 0)
Tail index 1/ξ4
End of the support μ − σ/ξlower limit −4
P(maximum > 6)0.02528
1-in-100 return level z_T8.633
Gumbel 1-in-100 level4.6
Ratio to the Gumbel level1.877
Probability of exceeding z_T in a year0.01
Worked examples
1. The worst of 250 days
Step 1 of 3
1
Exceedances of a 99% level
A daily loss beats its 99% quantile with probability 1%. Over 250 independent days, the expected number of such days is 2.5, and the chance of none at all is
P(M250≤q0.99)=0.99250=0.0811
2. Return levels for a catastrophe line
Step 1 of 3
1
The fitted distribution
The largest single catastrophe loss in a year for an insurer's property line follows a GEV with μ = €100m, σ = €40m and ξ = 0.2 (illustrative values).
3. Why return tails are Fréchet-type
Step 1 of 3
1
Fit both
Simulate 400 years of 250 daily losses, once normal and once Student t with ν = 4, take each year's maximum and fit a GEV to each set by maximum likelihood (M-D08: Maximum likelihood estimation).
4. A tail quantile from exceedances
Step 1 of 3
1
The fit
Of 2,500 daily losses, 50 exceed u = 2.5%, and a GPD fitted to the excesses has β = 0.6 and ξ = 0.25 (illustrative values). So P(loss > 2.5%) = 2%.
Common mistakes
Each line below is a tempting slip. Spot what is wrong, then check.
Reaction
Check what stuck
Question 1 · objective 1
Daily losses are independent with the same distribution. What is the probability that none of 20 days beats the 95% quantile? Three decimals.
Question 2 · objective 3
Yearly maxima follow a Gumbel distribution with μ = 10 and σ = 2. What is the 1-in-100-year return level?
Question 3 · objective 2
A fitted GEV has shape ξ = 0.25. What is the tail index α = 1/ξ?
Question 4 · objective 2
A GEV has μ = 5, σ = 1 and ξ = −0.25. What is the largest value it can take?
Question 5 · objective 4
Why would fitting a Gumbel (ξ = 0) to the yearly maxima of fat-tailed losses be risky?
Glossary
Block maximum
Block maximum: the largest value in a block of observations, such as the worst daily loss in a year or the largest claim in a year.
GEV distribution
Generalised extreme value (GEV) distribution: the only possible limit of suitably shifted and scaled block maxima. Parameters: location μ, scale σ, shape ξ.
Shape parameter ξ
Shape parameter ξ: sets the tail. ξ > 0 heavy (power) tail, ξ = 0 light (exponential) tail, ξ < 0 a finite upper end.
Gumbel distribution
Gumbel distribution: the GEV with ξ = 0. Maxima of light-tailed variables such as the normal end up here.
Fréchet case
Fréchet case: the GEV with ξ > 0. Maxima of heavy-tailed variables with power tails, such as the Student t, end up here.
Weibull case
Weibull case (reversed Weibull): the GEV with ξ < 0, which has a finite upper end point μ − σ/ξ.
Tail index
Tail index α = 1/ξ: for a power tail, P(X > x) falls roughly like x^(−α). Smaller α means a heavier tail; moments of order α and above are infinite.
Return level
Return level z_T: the level a yearly maximum exceeds with probability 1/T, the 'T-year' or '1-in-T' level.
Generalised Pareto distribution (GPD)
Generalised Pareto distribution (GPD): the limit distribution of excesses over a high threshold, P(Y > y) = (1 + ξy/β)^(−1/ξ).
Threshold
Threshold u: the level above which observations are treated as tail data in the peaks-over-threshold method.
Embrechts, Klüppelberg and Mikosch, Modelling Extremal Events for Insurance and Finance: the full theory of maximum domains of attraction and its uses in insurance. Advanced.
Origins
R. A. Fisher and L. H. C. Tippett, “Limiting forms of the frequency distribution of the largest or smallest member of a sample”, Mathematical Proceedings of the Cambridge Philosophical Society 24(2), 180–190, 1928.
Stuart Coles, An Introduction to Statistical Modeling of Extreme Values, Springer Series in Statistics, Springer, 2001 (chapter 3, classical extreme value theory and models; chapter 4, threshold models).
Paul Embrechts, Claudia Klüppelberg and Thomas Mikosch, Modelling Extremal Events for Insurance and Finance, Stochastic Modelling and Applied Probability vol. 33, Springer, 1997.
James Pickands III, “Statistical inference using extreme order statistics”, The Annals of Statistics 3(1), 119–131, 1975.
A. A. Balkema and L. de Haan, “Residual life time at great age”, The Annals of Probability 2(5), 792–804, 1974.