Guide détaillé à venir
Nous préparons un guide éducatif complet pour le Value at Risk (VaR). Revenez bientôt pour des explications étape par étape, des formules, des exemples concrets et des conseils d'experts.
Value at Risk (VaR) is a statistical measure that quantifies the potential financial loss a portfolio may experience over a defined time horizon at a given confidence level. For example, a 1-day 95% VaR of $1 million means that there is a 95% probability that the portfolio will not lose more than $1 million in a single trading day — or equivalently, a 5% chance that losses will exceed $1 million on any given day. VaR became the dominant risk metric in financial institutions following J.P. Morgan's publication of the RiskMetrics methodology in 1994 and was subsequently embedded in the Basel II and Basel III regulatory frameworks for bank capital adequacy. VaR can be calculated using three main approaches. The Historical Simulation method uses the actual historical distribution of portfolio returns, replaying past return observations to construct a loss distribution and reading off the relevant percentile. The Parametric (Variance-Covariance) method assumes returns are normally distributed and uses the portfolio's mean return, standard deviation, and the normal distribution's z-score to compute VaR analytically. The Monte Carlo Simulation method generates thousands of random return scenarios based on specified distributional assumptions and correlations, producing a synthetic loss distribution from which VaR is estimated. Each method has strengths and limitations. Historical simulation is intuitive but limited by the quality and length of the historical data set. The parametric method is computationally simple but fails when return distributions have fat tails or non-linear risk exposures (e.g., options). Monte Carlo is the most flexible but computationally intensive and sensitive to model assumptions. VaR is extensively used by commercial banks, investment banks, hedge funds, asset managers, insurance companies, and pension funds. It is a regulatory requirement for market risk capital calculations under Basel III (using Internal Models Approach) and is disclosed in annual reports of major financial institutions. Despite its ubiquity, VaR has well-documented limitations: it does not describe the severity of losses beyond the confidence threshold (tail losses), it can underestimate risk during market stress, and it can create false precision. Complementary measures such as Expected Shortfall (CVaR) and stress testing are always used alongside VaR for comprehensive risk management.
Parametric VaR = W × (μ × T − z × σ × √T) Historical VaR = Percentile(P&L distribution, 1 − confidence level)
- 1Define the parameters: choose the confidence level (typically 95% or 99%), the time horizon (1 day or 10 days), and the portfolio value.
- 2For Parametric VaR: collect the portfolio's daily return series, compute the mean daily return (μ) and standard deviation (σ), then calculate VaR = W × (z × σ × √T − μ × T), where z = 1.645 for 95% or 2.326 for 99%.
- 3For Historical Simulation: collect at least 250–500 daily return observations, sort them from worst to best, and identify the 5th percentile (for 95% VaR) or 1st percentile (for 99% VaR) of this historical loss distribution.
- 4For Monte Carlo: specify the return distribution parameters (mean, volatility, correlations), generate 10,000+ random return scenarios, construct the simulated P&L distribution, and read off the appropriate percentile.
- 5Scale from 1-day to longer horizons: multiply the 1-day VaR by √T to estimate T-day VaR (valid under the square-root-of-time rule, which assumes independent returns).
- 6Backtest the VaR model by counting how often actual losses exceeded the VaR estimate ('exceptions'). For 99% VaR, about 2.5 exceptions per year (out of 250 trading days) is expected; significantly more indicates model failure.
- 7Report VaR alongside Expected Shortfall (CVaR) — the average loss given that the loss exceeds VaR — to characterize the tail of the loss distribution beyond the VaR threshold.
There is a 5% chance of losing more than $19,374 on any given trading day.
Using the parametric formula: VaR = $1,000,000 × (1.645 × 0.012 − 0.0004) = $1,000,000 × (0.01974 − 0.0004) = $1,000,000 × 0.01934 = $19,340 (approximately $19,374 with precise z-score). This means that on 95% of trading days, the portfolio loss is expected to be below this amount. On 5% of days — roughly 12 to 13 trading days per year — losses are expected to exceed this figure. This VaR estimate assumes normally distributed returns and would underestimate risk if the return distribution has fat tails.
Basel III requires banks to hold capital against 10-day 99% VaR for market risk.
Under Basel III's Internal Models Approach, banks must compute 10-day 99% VaR for market risk capital requirements. Using the square-root-of-time rule: 10-Day VaR = 1-Day VaR × √10 = $500,000 × 3.162 = $1,581,139. The bank must hold regulatory capital of at least 3 times this VaR (the multiplier factor) as a capital buffer against market risk. This example illustrates how VaR directly drives regulatory capital requirements and balance sheet management decisions at major financial institutions.
Based on actual historical return distribution — no normality assumption needed.
For a hedge fund with 500 days of return history, the historical simulation method ranks all 500 daily P&L observations. The 25th worst daily loss (the 5th percentile = 500 × 0.05 = 25th observation) represents the 95% VaR. If the 25th worst day resulted in a 3.8% loss on a $2,000,000 portfolio, the historical VaR is $76,000. This method is particularly valuable for hedge fund portfolios with non-linear exposures (options, convertibles) where the normality assumption of the parametric method would fail, since it uses actual realized returns regardless of their distributional shape.
10.4% daily VaR — crypto's extreme volatility makes VaR very large relative to portfolio.
For a cryptocurrency portfolio with daily volatility of 4.5% (versus ~1% for typical equity portfolios), the 99% VaR is approximately $100,000 × (2.326 × 0.045 − 0.001) = $100,000 × 0.10367 = $10,377. This means there is a 1% chance of losing over 10% of the portfolio on any given day — a much higher absolute risk than conventional financial assets. Critically, the parametric model likely understates true crypto VaR due to fat-tailed, leptokurtic return distributions. Historical simulation or extreme value theory would typically produce higher estimates for such asset classes.
Basel III regulatory capital calculations for bank market risk under the Internal Models Approach, representing an important application area for the Value At Risk in professional and analytical contexts where accurate value at risk calculations directly support informed decision-making, strategic planning, and performance optimization
Daily risk limit monitoring for trading desks at investment banks and market-making firms, representing an important application area for the Value At Risk in professional and analytical contexts where accurate value at risk calculations directly support informed decision-making, strategic planning, and performance optimization
UCITS fund prospectus risk disclosures required by EU regulations (SRRI methodology), representing an important application area for the Value At Risk in professional and analytical contexts where accurate value at risk calculations directly support informed decision-making, strategic planning, and performance optimization
Portfolio optimization: constructing minimum-VaR portfolios within return targets, representing an important application area for the Value At Risk in professional and analytical contexts where accurate value at risk calculations directly support informed decision-making, strategic planning, and performance optimization
Insurance company risk-based capital calculations for investment portfolio market risk, representing an important application area for the Value At Risk in professional and analytical contexts where accurate value at risk calculations directly support informed decision-making, strategic planning, and performance optimization
{'case': 'Options Portfolios and Non-Linear VaR', 'description': 'For portfolios containing options, linear VaR models (parametric) are inappropriate because options have non-linear P&L profiles (delta, gamma, vega exposures). Full revaluation VaR (Historical Simulation or Monte Carlo) must be used, applying the return scenarios to reprice each option individually using its pricing model rather than assuming proportional P&L changes.'}
{'case': 'Negative Portfolio Value or Short Positions', 'description': 'For portfolios with net short positions, gains and losses are inverted relative to market movements. The VaR must be calculated from the perspective of the actual P&L distribution, not the return distribution of the underlying assets. Short positions benefit from market declines but lose from market advances, so VaR for short books must account for upside market risk rather than downside risk.'}
{'case': 'Correlation Breakdown During Stress', 'description': 'VaR models that rely on historical correlations between assets may dramatically understate portfolio VaR during market stress events, when previously uncorrelated assets become highly correlated (correlation goes to 1 in a crisis). Diversification benefits assumed in normal markets can disappear suddenly, causing actual losses to far exceed multi-asset VaR estimates.'}
| Confidence Level | Z-Score | Daily Exception Rate | Expected Exceptions/Year |
|---|---|---|---|
| 90% | 1.282 | 10% | ~25 days/year |
| 95% | 1.645 | 5% | ~12–13 days/year |
| 97.5% | 1.960 | 2.5% | ~6 days/year |
| 99% | 2.326 | 1% | ~2–3 days/year |
| 99.5% | 2.576 | 0.5% | ~1 day/year |
| 99.9% | 3.090 | 0.1% | ~1 day every 4 years |
What is the difference between VaR and Expected Shortfall (CVaR)?
VaR answers the question 'what is the maximum loss at a given confidence level?' — but it says nothing about how large losses can be beyond that threshold. Expected Shortfall (ES), also called Conditional VaR (CVaR), answers the complementary question 'if losses exceed VaR, what is the average loss?' ES is always greater than or equal to VaR and provides a more complete description of tail risk. For example, if 99% VaR is $1M, CVaR might be $1.8M — meaning losses beyond the VaR threshold average $1.8M. Basel III has progressively shifted from VaR toward Expected Shortfall as the primary regulatory risk measure due to CVaR's superior tail risk properties.
Why doesn't VaR tell you how bad the worst losses can be?
VaR is a quantile measure — it specifies a threshold that losses will not exceed with a given probability, but it provides no information about the distribution of losses beyond that threshold. Two portfolios can have identical VaRs but very different tail loss profiles: one might have moderate losses just beyond VaR, while another might experience catastrophic losses. This limitation is particularly severe for portfolios with options, credit derivatives, or other instruments with non-linear, highly skewed return distributions. The practical consequence is that VaR alone is insufficient for risk management and must be supplemented by Expected Shortfall and stress testing.
What confidence level should I use for VaR?
The most common choices are 95% (used in many internal risk management applications and by some regulators) and 99% (specified by Basel III for bank market risk capital). Some firms use 99.9% or higher for economic capital and stress testing purposes. The choice depends on the application: internal risk limits might use 95% VaR for daily monitoring (producing more frequent but smaller exceptions), while regulatory capital and risk tolerance frameworks typically use 99% VaR. Higher confidence levels produce larger VaR estimates and require more historical data for reliable estimation — 99.9% VaR from historical data requires at least 1,000 observations to have a meaningful number of tail events.
How is VaR backtested?
VaR models are backtested by comparing daily VaR estimates against actual subsequent P&L realizations and counting the number of 'exceptions' — days when actual losses exceeded the VaR estimate. For a correctly calibrated 99% VaR model, approximately 1% of days should produce exceptions — about 2 to 3 days per year out of 250 trading days. Basel III uses a traffic-light system: 0–4 exceptions per year is a green zone (model is acceptable), 5–9 is a yellow zone (model is suspect, additional capital charges apply), and 10+ is a red zone (model is rejected). Systematic over- or under-prediction of exceptions triggers regulatory scrutiny and model recalibration.
What are the main limitations of VaR?
VaR has several well-documented limitations. It can provide false precision — expressing risk as a single number can create misleading confidence. It does not measure tail risk (losses beyond the VaR threshold). Parametric VaR assumes normally distributed returns, which underestimates risk from fat-tailed distributions. Historical simulation is backward-looking and may not capture future market conditions. VaR can be procyclical — appearing low during calm markets when risk is actually accumulating, and spiking during crises when positions need to be cut. Additionally, during correlated stress events, diversification benefits assumed by VaR models can disappear, causing actual losses to far exceed VaR estimates.
How does the time horizon affect VaR?
VaR scales approximately with the square root of the time horizon under the assumption that daily returns are independent and identically distributed (i.i.d.). A 10-day VaR is approximately √10 ≈ 3.16 times the 1-day VaR. This scaling rule is embedded in Basel III (which specifies 10-day VaR for regulatory capital) and is mathematically valid when returns are uncorrelated across time. However, if returns exhibit serial autocorrelation — which many do at short horizons — the square-root-of-time rule overestimates VaR for positively autocorrelated assets and underestimates it for negatively autocorrelated (mean-reverting) ones.
What is VaR used for in practice by banks and investment firms?
Banks use VaR for regulatory capital calculations (Basel III market risk capital), daily risk limit monitoring for trading desks, counterparty credit exposure estimation, and economic capital allocation across business lines. Investment managers use VaR for portfolio construction (optimizing the portfolio to achieve target VaR within risk budget constraints), client reporting (disclosing portfolio risk in standardized form), fund prospectus risk disclosures (required in UCITS funds in the EU), and stress testing. Risk management systems at major banks compute VaR in real time across thousands of positions and asset classes to aggregate firm-wide market risk exposure.
Conseil Pro
Always complement VaR with Expected Shortfall (CVaR) and historical stress tests using actual crisis scenarios (2008, 2020). Ask: 'If we are in the 1% tail, how bad can it get?' — that is the question VaR cannot answer but risk management cannot ignore.
Le saviez-vous?
J.P. Morgan's 1994 RiskMetrics publication, which popularized VaR, was driven by a request from then-CEO Dennis Weatherstone who wanted a single daily report summarizing the firm's total risk in one number. The result — the '4:15 report' delivered 15 minutes after market close — changed global banking risk management forever.
Références