Value at Risk

Value at risk estimates a loss threshold over a stated horizon and confidence level, subject to the data, model, and liquidity assumptions used.

Value at risk (VaR) estimates a loss threshold for a portfolio or position over a stated time horizon and at a stated confidence level. A one-day 95% VaR of $2 million means the model places the 95th percentile of one-day loss at $2 million under its assumptions.

VaR is not the maximum possible loss. It does not say how large losses may be after the threshold is breached, and it is not a guarantee that breaches will occur at the modeled frequency.

Key Takeaways

  • A VaR number is incomplete without the portfolio scope, valuation time, horizon, confidence level, currency, and method.
  • VaR identifies a loss cutoff; Expected Shortfall estimates average loss beyond a cutoff.
  • Historical simulation, parametric methods, and Monte Carlo simulation can produce different results for the same portfolio.
  • Correlation, volatility, option behavior, liquidity, data history, and valuation models can materially change the estimate.
  • Backtesting checks model performance against observed outcomes, but stress testing remains necessary because historical data may not contain the relevant crisis.
  • VaR can support limits and risk reporting, but it should not be treated as a complete measure of Tail Risk or capital adequacy.

What a VaR Statement Means

Suppose a report states:

One-day 99% market VaR: $5 million.

Under the model and sign convention used, $5 million is the estimated 99th-percentile one-day loss. The estimate leaves approximately 1% of modeled outcomes beyond the cutoff. It does not describe the size of those worse outcomes.

The statement does not mean:

  • the portfolio cannot lose more than $5 million
  • a breach will happen exactly once every 100 trading days
  • a ten-day loss equals ten times the one-day VaR
  • the position can be liquidated within one day without affecting prices
  • all credit, liquidity, operational, or legal risk is included

Always read the methodology and scope before interpreting the amount.

Mathematical Definition

Let \(L\) be a loss variable, so larger positive values represent larger losses. At confidence level \(\alpha\), VaR can be written as:

$$ \operatorname{VaR}_{\alpha}(L) = \inf \left\{ \ell : \Pr(L \leq \ell) \geq \alpha \right\} $$

This is a quantile of the modeled loss distribution. For example, \(\alpha=0.95\) selects the 95th percentile.

Some systems model returns or profit and loss instead of positive losses. Their signs may be reversed. A report should state its convention rather than relying on the word “loss” alone.

Main VaR Methods

MethodHow it worksMain strengthsMain limitations
Historical simulationRevalues the current portfolio using historical changes in risk factorsPreserves many observed relationships and does not require normal returnsLimited by the selected history; unseen events and structural changes are absent
Parametric or variance-covarianceUses estimated volatility, correlation, and a chosen distribution or approximationFast and transparent for portfolios with relatively linear exposuresDistribution and linearity assumptions can understate options, jumps, and fat tails
Monte Carlo simulationGenerates many scenarios from a specified stochastic model and revalues the portfolioFlexible for nonlinear instruments and complex dependenciesComputationally intensive and only as credible as the simulation, parameters, and pricing models

Method labels are not enough. Two historical simulations can differ because of lookback window, observation weighting, data cleaning, valuation mapping, and treatment of missing prices. Two Monte Carlo models can differ because of distribution, dependence, volatility dynamics, and number of scenarios.

Worked Example: VaR and Expected Shortfall

Assume a risk model reports:

  • portfolio value: $100 million
  • horizon: one trading day
  • confidence level: 95%
  • VaR: $2 million
  • expected shortfall: $3.4 million

The VaR is the modeled loss cutoff. The expected shortfall estimates the average loss in the modeled tail beyond the selected confidence level. It would be incorrect to report $2 million as the worst possible loss or to assume that every breach will be close to $3.4 million.

The committee should also ask:

  • Which positions and risk factors are included?
  • Were options fully revalued or approximated?
  • Does the historical window contain stressed conditions?
  • Are illiquid positions marked with reliable prices?
  • How often did actual losses exceed VaR in backtesting?
  • What do stress scenarios show beyond the model’s ordinary range?

Horizon and Confidence Level

A higher confidence level generally produces a larger VaR because the cutoff moves farther into the loss tail. A longer horizon can also produce a larger estimate, but the relationship is not necessarily linear.

Square-root-of-time scaling is sometimes used:

$$ \operatorname{VaR}_{T} \approx \operatorname{VaR}_{1}\sqrt{T} $$

That approximation depends on strong assumptions, including stable, independent risk-factor changes and a portfolio that does not materially change. It can be misleading when volatility clusters, returns jump, positions are nonlinear, trading occurs during the horizon, or markets become illiquid.

Comparisons are meaningful only when horizon, confidence level, currency, valuation basis, and portfolio scope are aligned.

Diversification and Aggregation

Portfolio VaR can be lower than the sum of stand-alone VaRs when positions offset each other in the model. The apparent diversification benefit depends on estimated dependence and on whether relationships remain stable in stress.

Important checks include:

  • correlations across risk factors and business units
  • basis risk between positions assumed to offset
  • concentration in common issuers, sectors, currencies, or strategies
  • nonlinear exposure from options and structured products
  • aggregation across models with different data and horizons
  • whether diversification is restricted for limits or regulatory purposes

Diversification observed in calm periods can weaken when markets move sharply together.

Backtesting and Exceptions

Backtesting compares VaR forecasts with subsequently observed profit and loss. A day when the relevant loss exceeds VaR is often called an exception or breach.

Exception counts need context:

  • Too many exceptions may indicate underestimated risk, weak data, valuation errors, or a changed market regime.
  • Very few exceptions can reflect conservative estimates, but they do not prove the model captures tail severity.
  • Profit-and-loss definition matters: actual trading P&L, hypothetical P&L on unchanged positions, fees, intraday trading, and valuation adjustments may produce different comparisons.
  • Independence matters: clustered exceptions can reveal regime change even when the total count appears acceptable.

Backtesting is evidence about observed performance, not proof that the future distribution is known.

VaR, Stress Testing, and Risk Limits

VaR is most useful as one part of a broader framework:

ToolMain question
VaRWhat loss cutoff does the model estimate for a stated horizon and confidence level?
Expected shortfallWhat is the modeled average loss in the tail beyond the cutoff?
Stress testingWhat happens under a specified severe historical or hypothetical scenario?
Sensitivity analysisHow does value change when one or more risk factors move?
Position and concentration limitsHow much exposure may be held regardless of model output?
Liquidity analysisCan positions be exited or hedged within the assumed horizon and at what cost?

A limit should specify the measure, scope, calculation time, escalation threshold, owner, exceptions process, and response. Reducing reported VaR without reducing economic exposure can occur through model changes, stale inputs, omitted risks, or positions that exploit model weaknesses.

Historical Note: RiskMetrics

RiskMetrics was a published set of techniques and data for measuring market risk across fixed income, equities, foreign exchange, commodities, and derivatives. The 1996 technical document helped make variance-covariance VaR methods and market-risk data broadly accessible.

RiskMetrics is not a separate loss statistic from VaR. It is a named methodology and historical risk framework with specific volatility, correlation, mapping, and distribution assumptions. Later RiskMetrics 2006 work revised the volatility process and return-distribution treatment. An analyst should identify the exact methodology and version rather than using “RiskMetrics” as a generic synonym for all risk measurement.

Risks and Limitations

  • Tail severity: VaR does not measure how bad losses become beyond the cutoff.
  • Model risk: assumptions, data, coding, valuation, and use can be wrong. See Model Risk.
  • Historical dependence: a chosen lookback period may omit relevant crises or overemphasize one regime.
  • Liquidity: modeled prices may not be executable for the position size or horizon.
  • Nonlinearity: approximations can fail for options, structured products, and path-dependent exposures.
  • Changing positions: a static model may not represent trading, hedging, margin calls, or forced sales during stress.
  • Coverage: market VaR may omit credit migration, default, funding, legal, operational, and settlement risk.
  • Incentives: a single headline number can encourage risk taking just beyond the modeled cutoff.

Common Mistakes

  • Calling VaR a maximum loss.
  • Reporting a VaR amount without horizon and confidence level.
  • Comparing VaR estimates calculated with different methods or scopes.
  • Scaling mechanically across time without checking assumptions.
  • Treating low recent volatility as permanent.
  • Assuming modeled diversification will remain available in stress.
  • Ignoring valuation uncertainty and market liquidity.
  • Using backtesting as a substitute for scenario analysis.
  • Changing the model to reduce exceptions without addressing the underlying exposure.

Official Sources

These sources address specified bank market-risk and regulatory-capital contexts. A firm’s internal VaR may use different horizons, confidence levels, methods, and scope. Applicable regulatory requirements depend on the institution and jurisdiction.

  • Expected Shortfall: The average modeled loss beyond a selected confidence cutoff rather than the cutoff itself.
  • Tail Risk: The broader exposure to extreme outcomes, nonlinear losses, concentration, and model failure.
  • Stress Testing: Analysis under specified severe conditions that may not be represented in the VaR distribution.
  • Standard Deviation: A dispersion measure used by some parametric VaR methods but not equivalent to a loss quantile.
  • Model Risk: The adverse-decision risk created by assumptions, data, implementation, validation, or misuse of the VaR model.

FAQs

What does 95% VaR mean?

It is the modeled 95th-percentile loss over the stated horizon. Approximately 5% of modeled outcomes lie beyond that cutoff, but VaR does not show how severe those outcomes may be.

Can actual loss exceed VaR?

Yes. Losses beyond the VaR cutoff are expected under the model, and actual markets can behave differently from the model.

Is a lower VaR always better?

No. A lower result may reflect less exposure, but it may also result from a different method, shorter horizon, lower confidence level, stale data, assumed diversification, or omitted risk.

Educational Use

This article provides general financial education. It is not personalized investment, trading, banking, regulatory, model-validation, capital, liquidity, or risk-management advice.

Browse Risk Management