Explain VaR and Expected Shortfall: parametric/historical/Monte Carlo methods,…
- User
Explain VaR and Expected Shortfall: parametric/historical/Monte Carlo methods, backtesting, stress scenarios, liquidity horizons, model risk/limitations, and regulatory views; add a concise numeric example.
- Assistant
Value at Risk (VaR) and Expected Shortfall (ES)
VaR quantifies the maximum potential loss on a portfolio over a time horizon (e.g., 1-day, 10-day) at a confidence level (e.g., 99%), such that losses exceed VaR with probability 1-α. E.g., 99% 1-day VaR of 1M tomorrow.
ES (aka Conditional VaR or TailVaR) measures the average loss exceeding the VaR threshold, capturing tail risk magnitude. ES > VaR, providing a more coherent risk measure (subadditive).
Both are used for market, credit, and operational risk; ES preferred for tails.
Computation Methods
Method Description Pros Cons VaR Formula/Example ES Adaptation Parametric (Variance-Covariance) Assumes normal/elliptical returns; uses mean (μ), std dev (σ), correlations. Fast, analytical. Ignores fat tails, non-normality, volatility clustering. VaR_α = μ*t + Z_α * σ * √t (Z_α = inverse normal CDF, e.g., 2.326 for 99%). ES_α = μ*t + [φ(Z_α)/(1-α)] * σ * √t (φ=normal density). Historical Simulation Ranks empirical historical P&L (e.g., 5y daily returns); VaR=percentile. Non-parametric, captures non-normality. Assumes history repeats; "ghosting" in calm markets; data-hungry. Sort 1,000 obs; 99% VaR=10th worst loss. Average of losses beyond VaR percentile. Monte Carlo Simulate 10k+ paths via stochastic models (e.g., GARCH for vols, copulas for deps); rank P&L. Flexible (any dist, paths, illiquids); forward-looking. Computationally intensive; model risk. Percentile of simulated losses. Average of tail simulations. Hybrids (e.g., Filtered Historical: scale by current vol) common.
Backtesting
Validates model: Count "exceptions" (actual loss > VaR). For 99% VaR over 250 days, expect ~2.5 exceptions.
- Kupiec (Unconditional Coverage): Likelihood ratio test if observed exceptions ≈ expected (binomial).
- Christoffersen (Conditional Coverage): Adds independence (clusters bad).
- Evans-Sarima: Spectral for autocorrelation.
Regulatory: Basel "green/yellow/red" zones (0-4, 5-9, ≥10 exceptions over 250 days → model requalification).
Stress Scenarios
Apply predefined shocks (historical: e.g., 2008 crisis vols +50%; hypothetical: e.g., +200bp yield spike). Compute stressed VaR/ES.
- Reverse stress tests: Find scenarios causing breach.
- Complements VaR (ignores extremes).
Liquidity Horizons (Liquidity-Adjusted VaR)
Standard VaR assumes instant liquidation; adjust for illiquids:
- Scale horizon: VaR_T = VaR_1 * √(T/1) + bid-ask/impact costs, T=10-20 days for privates.
- Basel: 10-day min; multiply by √10 for 1-day models; rebalance every 3m.
- Multi-period: Simulate sequential liquidation.
Model Risk and Limitations
- VaR: Not subadditive (diversification can increase); ignores loss severity; procyclical (low vol → low VaR → more risk).
- ES: Better, but estimation error in tails; sensitive to VaR level.
- Common: Non-stationarity (regimes), jumps, path-dependence, correlations break in stress.
- Mitigants: Validation, sensitivity tests, multiple models, margins.
Regulatory Views (Basel III/FRTB)
- Market Risk: ES at 97.5% conf, 10-day horizon (VaR legacy). Internal Models Approach (IMA) needs ES backtest + P&L attribution.
- Std Approach: Sensitivity-based (δ-normalized buckets).
- Credit/Op Risk: VaR variants; ES emerging.
- Stress: Mandatory 3x stressed ES backtest.
- Liquidity: FRTB incorporates liquidity horizons explicitly.
- ES over VaR: Addresses LTCM 1998 failure (VaR blind to tails).
Concise Numeric Example
$10M equity portfolio; daily ret ~N(0%, 2%); 99% conf, 1-day.
- Parametric VaR: 2.326 × 0.02 × 10 = **465k w/ 1% prob.)
- ES: [0.3989/(0.01)] × 0.02 × 10 ≈ $796k. (Avg loss given >VaR: ~1.71×VaR.)
Historical: Suppose sorted losses: ... -500k, -900k, ... → VaR= -900k etc.)= -$750k.
Monte Carlo: Sim 10k paths → similar percentiles. Stress (+50% vol): VaR→$698k.
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.