Value at Risk (VaR) and Conditional VaR¶
Difficulty advanced
Overview¶
VaR measures the maximum expected loss over a given time horizon at a specified confidence level. It's the standard risk metric used by financial institutions.
Definition¶
VaR(α, t) = Maximum loss not exceeded with probability α over time t
Example: VaR(95%, 1-day) = $10,000
Means: 95% confidence that loss won't exceed $10,000 in one day
Or: 5% chance of losing more than $10,000
where:
αconfidence level (e.g. 0.95 or 0.99) ·ttime horizon (typically 1 day, 10 days, or 1 year). does: the threshold loss that won't be exceeded with probability α. Does not describe what happens when the threshold is breached — that's CVaR's job.
Methods¶
1. Parametric (Variance-Covariance)¶
VaR = V_p × (μ - z_α × σ) × √t
where:
V_p = Portfolio value
z_α = Z-score for confidence level (1.645 for 95%, 2.326 for 99%)
σ = Portfolio volatility
t = Time horizon
does: assumes returns are normally distributed — VaR is then just a quantile of that normal. Fast to compute, but underestimates tail risk because actual returns have fat tails (use Cornish-Fisher or historical simulation if accuracy in the tail matters).
2. Historical Simulation¶
does: non-parametric — just reads VaR off the empirical return distribution. Captures observed tail shape (skew, kurtosis) but is limited to events that actually happened in your sample. Add a stress overlay for events outside the sample.
Portfolio VaR¶
does: the parametric portfolio VaR — used for regulatory reporting, internal risk limits, and capital allocation across desks; the
√(w'Σw)factor is the portfolio volatility, so correlations enter directly and diversification benefits show up automatically as lower VaR.
Conditional VaR (Expected Shortfall)¶
where:
αconfidence level matching the VaR threshold ·E[· | ·]conditional expectation ·Lossportfolio loss random variable does: the expected loss conditional on a VaR breach — used in place of VaR for tail-sensitive limits, regulatory FRTB capital under the new market-risk framework, and any optimization where you want a sub-additive (coherent) risk measure.
Limitations of VaR¶
| Limitation | Issue | Solution |
|---|---|---|
| Not sub-additive | Portfolio VaR can exceed sum of parts | Use CVaR instead |
| Ignores tail shape | Doesn't capture tail severity | CVaR / Expected Shortfall |
| Normality assumption | Returns have fat tails | Historical or Monte Carlo |
| Static view | Assumes fixed portfolio | Dynamic/stress testing |
| Confidence choice | 95% vs 99% changes everything | Use multiple levels |
Practical Guidelines¶
- Use Multiple Methods — Compare parametric, historical, and MC
- Report CVaR Too — VaR alone understates tail risk
- Backtest Regularly — Validate VaR predictions
- Stress Test — VaR doesn't capture extreme events
- Short Horizon — 1-day and 10-day are standard
- Component VaR — Understand which positions contribute most
- Marginal VaR — Impact of adding a new position
Key Formulas Reference¶
Parametric VaR: V × z × σ × √t
Historical VaR: Percentile of historical returns
CVaR: E[Loss | Loss > VaR]
Portfolio VaR: √(w'Σw) × z × V × √t
Marginal VaR: ∂VaR/∂w_i = z × V × (Σw)_i / √(w'Σw)
where:
Vportfolio value ·zconfidence quantile ·σportfolio volatility ·thorizon ·wweights vector ·Σcovariance matrix ·(Σw)_iith element ofΣwdoes: compact reference for the canonical VaR variants and the marginal sensitivity to each weight — used in code, in risk reports, and in allocation decisions where Marginal VaR identifies which positions to trim first to reduce total VaR per unit notional.
Next Steps¶
- Position Sizing — Individual trade risk
- Stress Testing — Extreme scenario analysis
- Portfolio Optimization — Risk-aware allocation