Expected Shortfall (CVaR) and Tail Risk Measurement

Expected Shortfall (CVaR) and Tail Risk Measurement

Expected Shortfall calculation, advantages over VaR, regulatory adoption under FRTB, extreme value theory, and tail dependence modeling.

When to Activate

• User needs a coherent risk measure that captures tail losses beyond VaR
• Regulatory capital computation under FRTB (Fundamental Review of the Trading Book)
• Portfolio optimization with tail risk constraints
• Analyzing extreme loss scenarios and tail dependence between assets
• Comparing ES across portfolios for risk budgeting
• Stress testing that requires understanding the full tail distribution

Core Concepts

Expected Shortfall Definition

Expected Shortfall at confidence level alpha is the expected loss conditional on the loss exceeding VaR:

• ES_alpha = E[L | L > VaR_alpha]
• Also called Conditional VaR (CVaR), Tail VaR, or Average VaR
• For continuous distributions: ES_alpha = (1/(1-alpha)) * integral from alpha to 1 of VaR_u du
• ES always >= VaR at the same confidence level
• ES is a coherent risk measure (satisfies subadditivity); VaR is not

Coherent Risk Measures (Artzner et al.)

A risk measure rho is coherent if it satisfies:

1. Monotonicity: If X <= Y in all states, then rho(X) >= rho(Y)
2. Subadditivity: rho(X + Y) <= rho(X) + rho(Y) — diversification cannot increase risk
3. Positive homogeneity: rho(lambda * X) = lambda * rho(X) for lambda > 0
4. Translation invariance: rho(X + c) = rho(X) - c

VaR violates subadditivity for non-elliptical distributions. A merged portfolio can have higher VaR than the sum of individual VaRs. ES satisfies all four properties.

ES for Standard Distributions

Normal distribution: ES_alpha = sigma * phi(z_alpha) / (1 - alpha) - mu

• At 97.5%: ES = sigma * phi(1.96) / 0.025 = sigma * 2.338
• At 99%: ES = sigma * phi(2.326) / 0.01 = sigma * 2.665

Student-t distribution: ES is larger due to heavier tails

• ES_alpha = (nu + z_alpha^2) / (nu - 1) * f_t(z_alpha) / (1 - alpha) * sigma
• Lower degrees of freedom (nu) produce dramatically larger ES relative to VaR

Empirical (historical): ES = average of all losses beyond the VaR quantile

• If 99% VaR is the 5th worst of 500 observations, ES_99 = average of the 5 worst losses

Regulatory Shift: VaR to ES (FRTB)

Basel III.1 / FRTB replaces VaR with ES for internal models approach (IMA):

• ES computed at 97.5% confidence (calibrated to approximate 99% VaR under normality)
• Liquidity-adjusted ES: different holding periods for different risk factor categories
  • 10 days: large-cap equities, major FX, interest rates
  • 20 days: small-cap equities, corporate bonds, minor FX
  • 40 days: credit spreads, exotic FX
  • 60 days: certain credit and equity volatility factors
  • 120 days: structured credit, illiquid positions
• Aggregation: ES_liquidity-adjusted = sqrt(sum over j of (ES_j * sqrt(LH_j/10))^2) approximately
• Stressed ES calibrated to worst 12-month period (same concept as stressed VaR)
• P&L attribution test: compare risk-theoretical P&L to hypothetical P&L

Methodology

ES Estimation Methods

Parametric ES

1. Fit a distribution to return data (normal, Student-t, skewed-t)
2. Compute ES analytically using the closed-form expression for that distribution
3. Best for portfolios with approximately elliptical return distributions
4. Fast but misses complex tail behavior

Historical ES

1. Compute portfolio P&L for each historical scenario
2. Sort losses from worst to best
3. Identify the VaR threshold (e.g., 5th worst of 500 for 99%)
4. ES = average of all losses at or beyond the VaR threshold
5. Sample size matters: with 500 days at 99%, ES is the average of only 5 observations — high estimation uncertainty
6. Bootstrap confidence intervals around ES estimate

Monte Carlo ES

1. Simulate N scenarios from fitted model (e.g., 100,000 paths)
2. Revalue portfolio under each scenario
3. Sort simulated P&L, compute ES as average of worst (1-alpha)*N losses
4. More stable than historical due to larger sample in the tail
5. Model risk: results depend on assumed stochastic process

Extreme Value Theory (EVT)

EVT provides a principled framework for modeling the tail beyond available data:

Generalized Pareto Distribution (GPD)

• For exceedances over a high threshold u: F_u(x) = 1 - (1 + xi*x/beta)^(-1/xi)
• Shape parameter xi > 0: heavy tails (Pareto-like); xi = 0: exponential tails; xi < 0: bounded tail
• Financial returns typically have xi in range [0.1, 0.4] — heavy tails confirmed
• ES from GPD: ES = VaR/(1-xi) + (beta - xi*u)/(1-xi)

POT (Peaks Over Threshold) Method

1. Choose threshold u (typically 90th-95th percentile of losses)
2. Fit GPD to exceedances above u
3. Extrapolate to desired quantile and compute ES
4. Mean excess plot helps select threshold: should be approximately linear above u for GPD fit
5. Advantage: can estimate ES at confidence levels beyond available data (e.g., 99.9%)

Tail Dependence

Tail dependence measures the probability that two assets experience extreme losses simultaneously:

• Upper tail dependence: lambda_U = lim P(X > F_X^{-1}(u) | Y > F_Y^{-1}(u)) as u->1
• Normal copula: lambda_U = 0 for all rho < 1 — underestimates joint tail risk
• Student-t copula: lambda_U > 0 for rho > -1 and finite df — captures tail dependence
• Clayton copula: lower tail dependence (crashes), no upper tail dependence
• Gumbel copula: upper tail dependence, no lower tail dependence
• For portfolio ES, tail dependence structure matters more than linear correlation

ES Decomposition and Risk Budgeting

• Component ES: ES_i = w_i * partial(ES)/partial(w_i) — Euler allocation
• Sum of component ES equals total portfolio ES
• Risk contribution: RC_i = Component_ES_i / Total_ES
• Risk budgeting: set target risk contributions and solve for weights
• ES-based optimization: min ES subject to return constraint (linear program for empirical ES)

Examples

Historical ES Calculation

500 daily P&L observations for a $10M portfolio.
Sorted worst losses: -$620K, -$580K, -$540K, -$490K, -$470K, ...
99% VaR = $470K (5th worst).
99% ES = average(-$620K, -$580K, -$540K, -$490K, -$470K) = $540K.
ES exceeds VaR by $70K (15% higher), capturing the tail severity.

EVT-Based ES at 99.9%

Fit GPD to 50 exceedances above the 90th percentile threshold.
Estimated xi = 0.25, beta = $150K, threshold u = $300K.
99.9% VaR (GPD) = $300K + ($150K/0.25)*[(500*0.001/50)^(-0.25) - 1] = $892K.
99.9% ES = $892K/(1-0.25) + ($150K - 0.25*$300K)/(1-0.25) = $1,289K.

FRTB Liquidity-Adjusted ES

Equity desk ES(10-day): $2.1M
Credit desk ES(40-day): $1.8M
FX desk ES(10-day): $900K
Liquidity-adjusted ES = sqrt((2.1)^2 + (1.8*sqrt(40/10))^2 + (0.9)^2)
                       = sqrt(4.41 + 12.96 + 0.81) = sqrt(18.18) = $4.26M

Quality Gate

• ES estimates must include confidence intervals (bootstrap or analytical)
• Historical ES requires minimum 5 tail observations; prefer at least 10 for stability
• GPD fit must be validated: Q-Q plot, Anderson-Darling test on exceedances
• Threshold selection for POT must be justified (mean excess plot, stability of xi and beta)
• Tail dependence model must be tested against empirical tail dependence estimates
• ES backtesting: use multinomial test or bootstrap test (ES backtesting is harder than VaR backtesting)
• Regulatory ES must use correct liquidity horizons per FRTB risk factor categories
• Component ES must sum to total ES (verify Euler decomposition numerically)
• Compare parametric, historical, and MC estimates; investigate divergences
• Re-estimate EVT parameters at least quarterly