17.2 Value at Risk & Conditional Value at Risk

17.2 Value at Risk & Conditional Value at Risk

While downside dispersion metrics describe historical return variability, institutional risk management requires forward-looking probabilistic models to quantify capital-at-risk, establish regulatory capital buffers, and protect against catastrophic solvency failures. Value at Risk (VaR) has emerged as the universal institutional benchmark for expressing portfolio risk in a single, intuitive dollar or percentage figure. However, because VaR provides no insight into the severity of losses beyond its cutoff threshold, modern risk management frameworks pair VaR with Conditional VaR (Expected Shortfall), macro stress testing, and reverse stress testing.

                       Institutional Tail Risk Analytics Architecture
                                              │
        ┌─────────────────────────────────────┼─────────────────────────────────────┐
        │                                     │                                     │
Value at Risk (VaR) Models            Conditional VaR / ES                 Stress Testing & Scenarios
  • 95% & 99% Confidence Horizons       • Average Loss Beyond VaR             • Historical Crises (1987, 2008)
  • Parametric (Delta-Normal)           • Coherent Risk Measure               • Hypothetical Macro Shocks
  • Historical Simulation               • Subadditivity: ρ(X+Y) ≤ ρ(X)+ρ(Y)   • Sensitivity Factor Shocks
  • Monte Carlo Full Revaluation        • Tail Fatness Quantification         • Reverse Stress Testing

---

1. Value at Risk (VaR): Core Concepts & Parameters

Formal Definition and Mathematical Statement

Value at Risk (VaR) is defined as the maximum expected loss that a portfolio can experience over a specified time horizon ($T$) at a given confidence level ($\alpha$), under normal market conditions.

Equivalently, VaR represents the threshold loss such that the probability ($p = 1 - \alpha$) of experiencing a loss greater than or equal to VaR over horizon $T$ is exactly $p$:

$\text{Prob}(\text{Loss} > \text{VaR}_{\alpha}) = 1 - \alpha$

$\text{Prob}(\Delta V \le -\text{VaR}_{\alpha}) = 1 - \alpha$

   Value at Risk (VaR) Probability Density Function
     ▲
     │                     Normal Distribution Density
     │                            /---\
     │                           /     \
     │                          /       \
     │                         /    |    \
     │      Tail Loss Area    /     |     \
     │      (1 - α = 5%)     /      |      \
     │     ┌───────────────┐/       |       \
     │     │░░░░░░░░░░░░░░░/        |        \
     │    ─┴───────────────●────────┴────────┴───────► Portfolio Return / P&L
     │                -VaR (95%)   Mean (μ)
     │              (Cutoff Point)
     │◄── Tail Losses ────►│

The Three Core Components of a VaR Statement

Every meaningful VaR statement must explicitly state three parameters:

1. Holding Period / Time Horizon ($T$): The duration over which loss is measured (e.g., 1-day for commercial trading desks; 10-day for Basel banking regulation; 1-month or 1-year for pension funds and endowments).
2. Confidence Level ($\alpha$): Typically 95% ($z = 1.645$) or 99% ($z = 2.326$). A 95% 1-day VaR implies that on 95 out of 100 trading days (19 out of 20 days), losses will not exceed the VaR figure.
3. Monetary or Percentage Loss: The dollar amount (e.g., $2.5 million) or percentage of portfolio value (e.g., 3.2%) exposed to loss.

Example Statement: "Our $100 million fund has a 1-day 95% VaR of $1.65 million (1.65%)." This means there is a 5% probability that the fund will lose more than $1.65 million in a single trading day.

The Square-Root of Time Rule

Under the assumption that daily returns are independent and identically distributed (i.i.d.) with constant mean and variance, a 1-day VaR can be scaled to a $T$-day horizon using the Square-Root of Time Rule:

$\text{VaR}_{T\text{-day}} = \text{VaR}_{\text{1-day}} \times \sqrt{T}$

$\sigma_T = \sigma_{\text{daily}} \times \sqrt{T}$

Example: If a portfolio has a 1-day 95% VaR of $500,000, its 10-day 95% VaR under Basel regulatory standards is: $\text{VaR}_{\text{10-day}} = \$500,000 \times \sqrt{10} = \$500,000 \times 3.1623 = \mathbf{\$1,581,139}$

Cautionary Insight: The square-root rule breaks down when asset returns exhibit autocorrelation (trending or mean-reverting behavior) or volatility clustering (GARCH effects).

---

2. VaR Estimation Methodologies

Institutional practitioners compute VaR using three primary quantitative methodologies:

                           VaR Estimation Methodologies
       ┌───────────────────────────────┼───────────────────────────────┐
       │                               │                               │
Parametric (Delta-Normal)       Historical Simulation           Monte Carlo Simulation
  • Analytical closed-form        • Empirical distribution        • Stochastic factor paths
  • Assumes Normal Gaussian       • Replays historical shifts     • Full non-linear revaluation
  • VaR = -[μ - z·σ] · V_p        • Captures fat tails & skew     • Handles complex derivatives
  • Fast, but misses fat tails    • Backward looking / static     • Flexible, but compute heavy

1. Parametric (Variance-Covariance / Delta-Normal) Method

The Parametric VaR method assumes that portfolio returns follow a standard normal distribution described entirely by mean ($\mu_p$) and standard deviation ($\sigma_p$).

Mathematical Formula

$\text{Percentage VaR}_{\alpha} = -(\mu_p - z_{\alpha} \sigma_p) = z_{\alpha} \sigma_p - \mu_p$

$\text{Dollar VaR}_{\alpha} = \left( z_{\alpha} \sigma_p - \mu_p \right) \times V_p$

Where:

• $V_p$ = Total portfolio market value
• $\mu_p$ = Expected portfolio return over the horizon (frequently assumed to be 0 for short daily horizons)
• $\sigma_p$ = Portfolio standard deviation over the horizon
• $z_{\alpha}$ = Standard normal critical value for one-tailed confidence level $\alpha$:
  • 90% Confidence ($\alpha = 0.90$): $z = 1.282$
  • 95% Confidence ($\alpha = 0.95$): $z = 1.645$
  • 99% Confidence ($\alpha = 0.99$): $z = 2.326$

Worked Example: Multi-Asset Parametric VaR

A wealth manager holds a $50,000,000 portfolio with an expected daily return of 0.02% and a daily volatility of 1.20%. Calculate the 1-day 95% VaR and 1-day 99% VaR in dollar terms:

1. 1-Day 95% VaR ($z = 1.645$): $\text{VaR}_{95\%} = [(1.645 \times 0.0120) - 0.0002] \times \$50,000,000 = (0.01974 - 0.0002) \times \$50,000,000 = 0.01954 \times \$50,000,000 = \mathbf{\$977,000}$
2. 1-Day 99% VaR ($z = 2.326$): $\text{VaR}_{99\%} = [(2.326 \times 0.0120) - 0.0002] \times \$50,000,000 = (0.027912 - 0.0002) \times \$50,000,000 = 0.027712 \times \$50,000,000 = \mathbf{\$1,385,600}$

2. Historical Simulation Method

The Historical Simulation method is a non-parametric approach that requires no assumptions about the shape of the return distribution. It takes current portfolio holdings and re-evaluates their hypothetical performance across a historical window of actual market returns (e.g., the last 1,000 trading days / 4 years).

Operational Mechanics

1. Collect historical daily price changes for all portfolio assets over $N$ days (e.g., $N = 1,000$).
2. Apply each historical day's percentage return vector to the current portfolio asset weights to generate 1,000 simulated daily portfolio P&L outcomes.
3. Rank-order the 1,000 simulated returns from worst to best.
4. The 95% VaR is the 50th worst return (the 5th percentile: $1,000 \times 0.05$).
5. The 99% VaR is the 10th worst return (the 1st percentile: $1,000 \times 0.01$).

3. Monte Carlo Simulation Method

The Monte Carlo Simulation method models portfolio risk by generating thousands of pseudo-random market factor trajectories using stochastic differential equations (e.g., Geometric Brownian Motion, jump-diffusion processes, or student-t copulas).

Operational Mechanics

1. Specify statistical distributions, drift rates, volatilities, and correlation matrices for all underlying risk factors.
2. Run 10,000 to 100,000 simulated draws of correlated factor paths.
3. Revalue the entire portfolio—including full non-linear revaluation of options, swaptions, and mortgage-backed securities—under each simulated trial.
4. Construct the simulated terminal P&L distribution and extract the desired percentile cutoff.

Comprehensive Comparison Matrix: VaR Methodologies

---

3. Conditional VaR (CVaR) / Expected Shortfall (ES)

The "Tail-Blindness" Flaw of Value at Risk

Despite its widespread regulatory adoption, VaR suffers from a fatal theoretical limitation: it is completely blind to the magnitude of losses beyond the confidence cutoff.

   VaR Tail-Blindness: Portfolio A vs. Portfolio B (Both have 95% VaR = $10M)
     ▲
     │     Portfolio A (Normal Tail)             Portfolio B (Fat Catastrophic Tail)
     │              /---\\                                     /---\
     │             /     \\                                   /     \
     │            /       \\                                 /       \
     │     ______/         \                         ______/         \
     │    /     │           \                       /   │   │         \
     └────┴─────●───────────┴──────► Loss     ─────┴────┴───●─────────┴──────► Loss
             -VaR ($10M)                            -CVaR  -VaR ($10M)
             Avg Tail Loss = $12M                   ($25M) Avg Tail Loss = $25M

Both portfolios in the diagram above have an identical 95% 1-day VaR of $10 million. However, when an extreme 5% tail event occurs:

• Portfolio A suffers an average loss of $12 million.
• Portfolio B (holding short credit default swaps or deep out-of-the-money options) suffers an average loss of $25 million.

VaR treats both portfolios as having identical risk, masking Portfolio B's catastrophic solvency vulnerability.

Mathematical Formulation of CVaR / Expected Shortfall

Conditional Value at Risk (CVaR)—also known as Expected Shortfall (ES) or Average Value at Risk—measures the expected monetary or percentage loss conditional on the loss exceeding the VaR threshold:

$\text{CVaR}_{\alpha} = E[\text{Loss} \mid \text{Loss} > \text{VaR}_{\alpha}]$

For a continuous loss distribution with density $f(x)$:

$\text{CVaR}_{\alpha} = \frac{1}{1 - \alpha} \int_{\alpha}^1 \text{VaR}_u \, du$

For a discrete historical or Monte Carlo simulation with $M$ tail observations exceeding VaR:

$\text{CVaR}_{\alpha} = \frac{1}{M} \sum_{i=1}^M L_i \quad \text{for all } L_i > \text{VaR}_{\alpha}$

The Axioms of Coherent Risk Measures

In 1999, Artzner, Delbaen, Eber, and Heath established the mathematical foundation for modern risk metrics, defining four essential axioms that a coherent risk measure $\rho(\cdot)$ must satisfy:

1. Monotonicity: If portfolio $X$ always generates worse losses than portfolio $Y$ in every state of the world ($X \le Y$), then $\rho(X) \ge \rho(Y)$.
2. Translation Invariance: Adding cash amount $c$ to a portfolio reduces its risk by exactly $c$: $\rho(X + c) = \rho(X) - c$.
3. Positive Homogeneity: Scaling a portfolio by factor $\lambda > 0$ scales its risk proportionately: $\rho(\lambda X) = \lambda \rho(X)$.
4. Subadditivity: The risk of a combined portfolio cannot exceed the sum of individual portfolio risks: $\rho(X + Y) \le \rho(X) + \rho(Y)$

                     Subadditivity: The Diversification Axiom

         Standalone Portfolio X:                Standalone Portfolio Y:
         Risk = ρ(X)                            Risk = ρ(Y)
              │                                      │
              └───────────────────┬──────────────────┘
                                  │
                                  ▼
                      Combined Portfolio (X + Y)
                      Risk = ρ(X + Y)
                                  │
                     ┌────────────┴────────────┐
                     ▼                         ▼
         If ρ(X + Y) ≤ ρ(X) + ρ(Y)     If ρ(X + Y) > ρ(X) + ρ(Y)
         ═════════════════════════     ═════════════════════════
         SUBADDITIVITY HOLDS           SUBADDITIVITY FAILS (VIOLATION)
         • Diversification is rewarded • Merging increases measured risk
         • Coherent (CVaR / ES)        • Non-Coherent (VaR in non-normal)

Why VaR Fails Subadditivity

• VaR is NOT a coherent risk measure because it fails the subadditivity condition when returns are non-normal. For digital options, short credit exposures, or highly skewed assets, combining two portfolios can produce a VaR that is greater than the sum of the individual VaRs (${\text{VaR}(X+Y) > \text{VaR}(X) + \text{VaR}(Y)}$), penalizing diversification.
• CVaR / Expected Shortfall is strictly COHERENT across all distributions, always satisfying subadditivity and accurately reflecting diversification benefits.

---