# EVT Peaks-Over-Threshold (GPD) VaR & ES Formula Guide

> Derive Generalized Pareto Distribution (GPD) Peaks-Over-Threshold VaR and Expected Shortfall formulas, mean excess plots, and tail index bounds.

Canonical: https://quantcalctools.com/guides/extreme-value-theory-peaks-over-threshold-gpd-tail-es-model
Last updated: 2026-11-14T22:21:42.699Z

**Publication:** QuantCalc Tools
**Category:** Parametric Var Expected Shortfall
**Published:** 2026-11-14T12:55:25.915Z
**Author:** Elena Rostova, FRM, CQF (FRM, Certificate in Quantitative Finance (CQF))
**Reviewer:** Not recorded

## Answer-first summary

By the Pickands-Balkema-de Haan theorem, tail losses exceeding a high threshold converge to a Generalized Pareto Distribution, enabling closed-form 99.9% VaR and Expected Shortfall extrapolation.

## Key takeaways

- Pickands-Balkema-de Haan theorem proves excess losses above a high threshold u converge to a Generalized Pareto Distribution (GPD), decoupling tail calibration from quiet central returns.
- Closed-form GPD Expected Shortfall is finite only when tail shape xi < 1.0, with the extreme ratio ES_alpha / VaR_alpha converging asymptotically to 1 / (1 - xi).
- At xi = 0.28 (typical equity index crash tail), 99.9% GPD Expected Shortfall ($9.03M) is 2.42x larger than Gaussian 99.9% ES ($3.73M) on an identical $100M portfolio.
- Never fit GPD to raw clustered returns; pre-whiten via GJR-GARCH first so exceedances above the 95th-percentile threshold satisfy independent arrival assumptions.

## Related financial entities

- **Basel III FRTB 97.5% Expected Shortfall (ES) Tail Metric** (FRTB-ES-975) — license: Basel III FRTB 97.5% Expected Shortfall (ES) Tail Metric (FRTB-ES-975) refers to an institutional financial benchmark and regulatory standard monitored by BIS Basel Committee MAR33 (INTL), currently calibrated at 97.5% (Confidence Interval) as of October 2026.
- **Basel III FRTB 97.5% Expected Shortfall (ES) Tail Metric — 2026 Standard** (FRTB-ES-975-STD) — license: FRTB-ES-975 2026 Institutional Benchmark Framework is defined as the standardized quantitative and compliance methodology published by BIS Basel Committee MAR33 governing Basel III FRTB 97.5% Expected Shortfall (ES) Tail Metric across INTL capital markets (97.5% Confidence Interval, October 2026).
- **Cornish-Fisher Skew & Kurtosis Modified Parametric VaR** (CORNISH-FISHER-VAR) — product: Cornish-Fisher Skew & Kurtosis Modified Parametric VaR (CORNISH-FISHER-VAR) refers to an institutional financial benchmark and regulatory standard monitored by GARP FRM Quantitative Standard (INTL), currently calibrated at 99.0% (Non-Normal Quantile) as of October 2026.

## Analysis

Fitting a distribution to 100% of daily portfolio returns forces the 95% quiet-market center to dictate the 0.5% catastrophic crash tail—a compromise that fails whenever flash crashes, currency de-pegs, or sovereign defaults obey power-law physics distinct from everyday Gaussian noise. **Extreme Value Theory (EVT)** via the **Peaks-Over-Threshold (POT)** method discards the noisy interior and invokes the **Pickands-Balkema-de Haan (1974/1975) theorem**: for almost any continuous loss distribution $F(L)$, excess losses $Y = L - u$ above a sufficiently high threshold $u$ converge asymptotically to a two-parameter **Generalized Pareto Distribution (GPD)** $G_{\xi, \beta}(y)$. Compare standard normal outputs in our [Value at Risk (VaR) calculator](/calculators/value-at-risk-calculator) and anchor this tail model in our [Parametric VaR & Expected Shortfall calibration guide](/guides/parametric-var-expected-shortfall-fat-tail-calibration-guide).

## Exact Closed-Form GPD Tail VaR and Expected Shortfall Equations

For positive loss convention $L$, select threshold $u$ (typically the 90th or 95th empirical loss quantile) where $N_u$ out of $N$ total observations exceed $u$. The GPD cumulative distribution for excess loss $y = L - u \ge 0$ is $G_{\xi, \beta}(y) = 1 - \left(1 + \frac{\xi y}{\beta}\right)^{-1/\xi}$ for tail shape index $\xi \ne 0$ (and $1 - e^{-y/\beta}$ for $\xi = 0$), with scale $\beta > 0$. Setting unconditional tail survival $\bar{F}(u + y) = \frac{N_u}{N}\big(1 - G_{\xi, \beta}(y)\big) = 1 - \alpha$ for any $\alpha > 1 - N_u/N$ and inverting yields closed-form **EVT-POT VaR and Expected Shortfall** (valid for $\xi < 1$):

$$VaR_\alpha = u + \frac{\beta}{\xi}\left[\left(\frac{N}{N_u}(1 - \alpha)\right)^{-\xi} - 1\right], \qquad ES_\alpha = \frac{VaR_\alpha}{1 - \xi} + \frac{\beta - \xi u}{1 - \xi}$$

## Worked Pickands-Balkema-de Haan GPD Tail Calibration (99.5% ES)

```formula
Worked EVT Peaks-Over-Threshold 97.5% ES & 99.9% Stress VaR ($100M Portfolio, N = 1,000 Days)
Threshold Selection: 95.0th Percentile Loss u = 1.820% ($1,820,000), Exceedances N_u = 50 (N_u / N = 0.05)
MLE Fit on 50 Excess Losses (y_i = L_i - 1.820%):
 Tail Shape Index (xi) = +0.280 (Fréchet heavy power-law tail; tail index alpha_tail = 1/0.28 = 3.57!)
 Scale Parameter (beta)= 0.640% ($640,000)

1) 97.5% EVT VaR & FRTB Expected Shortfall (1 - alpha = 0.025):
 Probability Ratio = (1000 / 50) * 0.025 = 0.5000
 VaR_97.5 = 1.820% + (0.640% / 0.28) * [ (0.5000)^(-0.28) - 1 ]
 = 1.820% + 2.2857% * (1.21419 - 1) = 1.820% + 0.4896% = 2.3096% ($2,309,588)
 ES_97.5 = [ 2.3096% / (1 - 0.28) ] + [ (0.640% - 0.28 * 1.820%) / (1 - 0.28) ]
 = 3.2078% + (0.1304% / 0.72) = 3.2078% + 0.1811% = 3.3889% ($3,388,889!)

2) 99.9% CCP Default-Fund Extrapolated VaR & ES (1 - alpha = 0.001 — Only 1-in-1,000 day event!):
 Probability Ratio = 20 * 0.001 = 0.0200
 VaR_99.9 = 1.820% + 2.2857% * [ (0.0200)^(-0.28) - 1 ]
 = 1.820% + 2.2857% * (2.98969 - 1) = 1.820% + 4.5479% = 6.3679% ($6,367,863)
 ES_99.9 = (6.3679% + 0.1304%) / 0.72 = 9.0254% ($9,025,365 — vs Gaussian 99.9% ES of ~3.73%!)
```

Observe the asymptotic ratio $\lim_{\alpha \to 1} \frac{ES_\alpha}{VaR_\alpha} = \frac{1}{1 - \xi}$! Under a Gaussian distribution ($\xi = 0$), $ES_\alpha / VaR_\alpha \to 1.00$ in the extreme tail—meaning Expected Shortfall barely exceeds VaR at $99.99\%$. Under a Fréchet power-law tail with $\xi = 0.28$, $ES_\alpha$ asymptotically exceeds $VaR_\alpha$ by $\frac{1}{0.72} = 1.3889\times$ ($+38.89\%$) forever, and if $\xi \ge 1.0$, Expected Shortfall is mathematically **infinite**!

![Extreme Value Theory Peaks-Over-Threshold calibration workflow from GARCH pre-whitening to GPD MLE](/api/figure?type=flow&title=McNeil-Frey%20Pre-Whitened%20EVT%20Peaks-Over-Threshold%20Pipeline&subtitle=Stripping%20GARCH%20volatility%20clustering%20first%20so%20excess%20losses%20satisfy%20Pickands-Balkema-de%20Haan%20independence.&steps=1.%20GARCH%20Pre-Whiten%3A%3AFilter%20r_t%20into%20i.i.d.%20standardized%20losses%20z_t%7C2.%20Mean%20Excess%20Plot%3A%3ASelect%20lowest%20threshold%20u%20where%20e%28u%29%20turns%20upward-linear%7C3.%20Fit%20GPD%20%28xi%2C%20beta%29%3A%3AMaximize%20Grimshaw%20profile%20log-likelihood%20on%20N_u%20%3E%3D%2050%7C4.%20Extrapolate%20Tail%3A%3AEvaluate%20closed-form%20VaR_alpha%20%26%20ES_alpha%20for%20xi%20%3C%201&metrics=TYPICAL%20EQUITY%20XI%3A%3A%2B0.20..0.35%3A%3AFr%C3%A9chet%20heavy%20power-law%20class%7CASYMPTOTIC%20ES%2FVAR%3A%3A1%20%2F%20%281%20-%20xi%29%3A%3A1.39x%20ratio%20at%20xi%20%3D%200.28%7CFINITE%20ES%20BOUND%3A%3Axi%20%3C%201.00%3A%3AFirst%20moment%20diverges%20if%20xi%20%3E%3D%201%7CMIN%20EXCEEDANCES%20N_U%3A%3A40%20to%2080%3A%3ABias-variance%20threshold%20sweet%20spot)

## Choosing Threshold $u$: Sample Mean Excess Plots and Hill Estimator Pitfalls

Selecting threshold $u$ involves a classic bias-variance trade-off: setting $u$ too low (say, the 75th percentile) drags body observations into the tail and violates asymptotic GPD convergence (high bias); setting $u$ too high (say, the 99th percentile on 1,000 days, leaving $N_u = 10$ points) starves the MLE of data and explodes standard errors of $\hat{\xi}$ (high variance). Quantitative risk analysts inspect the **Sample Mean Excess Function** $e(u) = \mathbb{E}[L - u \mid L > u]$:
$$e(u) = \frac{\beta + \xi u}{1 - \xi}$$
Because $e(u)$ is an **exact linear function of $u$** with slope $\frac{\xi}{1 - \xi}$ once inside the true Pareto tail:
- **Upward-sloping linear $e(u)$** confirms a Fréchet heavy tail ($\xi > 0$).
- **Flat horizontal $e(u)$** signals exponential tail decay ($\xi = 0$, Gumbel domain, including Gaussian in limit).
- **Downward-sloping $e(u)$** signals bounded short-tail losses ($\xi < 0$, Weibull domain, such as long-only unlevered principal bounds).

## Tail Extrapolation Behavior Across Shape Parameter Regimes (Xi > 0, = 0, < 0)

| Asset Class / Strategy Tail | Typical GPD Shape $\hat{\xi}$ | Implied Tail Moment Index $1/\xi$ | Asymptotic $ES/VaR$ Ratio | Moments Finite Up To |
| :--- | :--- | :--- | :--- | :--- |
| Short-Dated G10 T-Bills | $0.05$ to $0.12$ | $8.3$ to $20.0$ | $1.05\times$ – $1.14\times$ | Skewness & Kurtosis finite |
| S&P 500 / FTSE 100 Daily Index | $0.22$ to $0.30$ | $3.3$ to $4.5$ | $1.28\times$ – $1.43\times$ | Variance finite; 4th moment diverges! |
| EM Sovereign Credit / FX Pegs | $0.32$ to $0.45$ | $2.2$ to $3.1$ | $1.47\times$ – $1.82\times$ | Variance finite; Skewness diverges! |
| Short Naked Weekly Options | $0.45$ to $0.65$ | $1.5$ to $2.2$ | $1.82\times$ – $2.86\times$ | Mean ES finite; **Variance infinite!** |
| Gumbel / Gaussian Limit | $0.00$ | $\infty$ | $\to 1.00\times$ | All polynomial moments finite |

![Comparison of 99.9% extrapolated stress Expected Shortfall across Gaussian, Student-t, Historical Simulation, and EVT GPD](/api/figure?type=bars&title=99.9%25%20Extreme%20Stress%20Expected%20Shortfall%20%28%24100M%20Portfolio%2C%20%24M%29&subtitle=Demonstrating%20why%20CCPs%20and%20Pillar%202%20Economic%20Capital%20models%20rely%20on%20EVT%20GPD%20beyond%2099.0%25%20confidence.&steps=Gaussian%20Parametric%3A%3A3.367%20sigma%20normal%20tail%20drastically%20understates%20crash%7C1000-Day%20Unfiltered%20HS%3A%3ATruncated%20at%20single%20worst%20historical%20observation%7CStudent-t%20%28nu%20%3D%205%29%3A%3ASymmetric%20fit%20constrained%20by%20central%2095%25%20returns%7CPOT%20GPD%20%28xi%20%3D%200.28%29%3A%3AUnconstrained%20Fr%C3%A9chet%20extrapolation%20beyond%2095%25%20u&metrics=GAUSSIAN%2099.9%25%20ES%3A%3A%243.73M%3A%3AAssumes%20exponential%20e%5E%28-z%5E2%2F2%29%20decay%7CRAW%201000D%20HS%2099.9%25%3A%3A%245.85M%3A%3ASingle%20noisy%20max%20loss%20point%7CSTUDENT-T%2099.9%25%20ES%3A%3A%247.41M%3A%3ASymmetric%20degrees%20of%20freedom%7CEVT%20GPD%2099.9%25%20ES%3A%3A%249.03M%3A%3ACaptures%20asymmetric%20left-tail%20xi%3D0.28)

## Critical Rule: Always Pre-Whiten With GARCH Before Fitting GPD!

Applying POT directly to raw clustered returns violates independence and biases $\hat{\xi}$ upward because a single 2-week volatility storm dumps 8 correlated points above $u$. Always filter returns via GJR-GARCH first—as shown in [Filtered Historical Simulation (FHS)](/guides/filtered-historical-simulation-garch-standardized-residual-var)—and fit GPD strictly to standardized innovations $\hat{z}_t$! Compare symmetric power tails in [Student-t Expected Shortfall scaling](/guides/student-t-expected-shortfall-degrees-of-freedom-tail-scaling), align internal capital models with [Basel III FRTB 97.5% Expected Shortfall](/entities/frtb-es-975-basel-iii-frtb-97-5-expected-shortfall-es-tail-metric) under [UK PRA](/regulator-index/prudential-regulation-authority-uk-173) Pillar 2 ICAAP supervision, and browse the [Parametric VaR & Expected Shortfall Topic Hub](/topics/parametric-var-expected-shortfall).

## Frequently asked questions

### Why is Peaks-Over-Threshold (POT) preferred over Block Maxima (GEV) in financial VaR modeling?

Block Maxima retains only the single worst loss per quarter or year, throwing away 98% of crisis data when multiple crashes occur inside the same quarter. POT uses every standardized exceedance above threshold u, slashing parameter standard errors.

### How do you read a Sample Mean Excess plot to pick threshold u for GPD fitting?

Plot average excess loss E[L - u | L > u] against candidate thresholds u from the 80th to 98th percentiles. Choose the lowest threshold u above which the graph stabilizes into an approximately straight upward-sloping line, typically leaving 40 to 80 exceedances.

### What does a GPD shape parameter xi = 0.30 imply about portfolio return moments?

The reciprocal 1 / xi = 3.33 equals the Pareto tail index m, meaning moments of order k >= 3.33 are infinite! Thus variance (k = 2) and skewness (k = 3) are finite, whereas theoretical fourth-moment kurtosis (k = 4) is infinite.

### Can Generalized Pareto VaR be smaller than threshold u?

Never when properly Domain-gated: GPD tail formulas apply strictly to confidence levels alpha > 1 - (N_u / N). For example, if u is set at the 95th loss percentile (N_u / N = 0.05), GPD evaluates quantiles at 97.5%, 99%, and 99.9%, never 90%.


## Source URLs

- https://www.jstor.org/stable/2240539
- https://www.sciencedirect.com/science/article/abs/pii/S0927539800000128
- https://www.bis.org/bcbs/publ/wp19.pdf

## Citation

- APA: Elena Rostova, FRM, CQF. (2026). Extreme Value Theory (EVT) Peaks-Over-Threshold: GPD VaR & ES Formula Guide. QuantCalc Tools. https://quantcalctools.com/guides/extreme-value-theory-peaks-over-threshold-gpd-tail-es-model
- Canonical article: [Extreme Value Theory (EVT) Peaks-Over-Threshold: GPD VaR & ES Formula Guide](https://quantcalctools.com/guides/extreme-value-theory-peaks-over-threshold-gpd-tail-es-model)
- Site research index: [QuantCalc Tools](https://quantcalctools.com/llms-full.txt)
