Value at Risk (VaR): Formula, Methods, and Examples

Value at Risk (VaR) is the most widely used risk metric in institutional finance. It distills a portfolio’s downside exposure into a single number — answering the question every investor and risk manager asks: “How much could I lose?” Popularized by JPMorgan’s RiskMetrics framework in 1994 and embedded in Basel banking regulations, VaR is now standard practice at banks, hedge funds, and asset managers worldwide. This guide covers what VaR measures, the three main calculation methods, how to interpret VaR numbers, and where VaR falls short. For video walkthroughs of these concepts, see the Portfolio Analytics & Risk Management course.

What is Value at Risk?

Value at Risk (VaR) is the loss threshold that is not expected to be exceeded over a given time horizon at a specified confidence level. It answers: “What is my loss threshold under normal market conditions, at a given level of confidence?”

VaR is defined by three parameters:

• Confidence level — the probability that losses will remain below the VaR threshold (e.g., 95% or 99%)
• Time horizon — the period over which losses are measured (e.g., 1 day, 10 days, 1 month)
• Portfolio value — the total dollar amount at risk

VaR can be expressed in dollars or as a percentage of portfolio value. A $50,000 VaR on a $1,000,000 portfolio is equivalent to a 5% VaR. By convention, VaR is reported as a positive number representing a loss — a VaR of $15,000 means a potential loss of $15,000.

It is important to distinguish VaR from standard deviation. Standard deviation measures the overall dispersion of returns in both directions. VaR focuses specifically on the left tail — the downside quantile at a chosen confidence level. Two portfolios with identical standard deviation but different skewness can have very different VaR numbers.

How to Interpret Value at Risk

The interpretation of VaR depends on the confidence level chosen. Higher confidence levels produce larger VaR numbers because they capture more extreme tail events:

These are one-tailed z-scores because VaR focuses exclusively on the left tail — losses. Using a two-tailed z-score (such as 1.96 for 95%) is a common error that overstates VaR.

The choice of time horizon depends on the use case. Trading desks typically use 1-day VaR for daily limit monitoring. Basel regulatory capital historically used a 10-day, 99% VaR. Portfolio managers and institutional investors often use monthly or annual VaR for strategic risk assessment.

To convert VaR between time horizons, use the square-root-of-time rule:

For example, 10-day VaR ≈ 1-day VaR × √10 ≈ 1-day VaR × 3.162. This approximation assumes returns are independent and identically distributed (i.i.d.) with stable volatility — conditions that hold reasonably well over short horizons but break down over longer periods.

The Value at Risk Formula

There are three standard approaches to calculating VaR. Each answers the same question — “What is my loss threshold at a given confidence level?” — but uses different assumptions and data inputs.

Parametric VaR (Variance-Covariance Method)

Parametric VaR assumes portfolio returns follow a normal distribution and calculates VaR directly from the portfolio’s standard deviation:

Where:

• P — portfolio value
• z — one-tailed z-score for the desired confidence level (1.645 for 95%, 2.326 for 99%)
• σ — portfolio standard deviation over the VaR time horizon

For multi-asset portfolios, computing σ_p requires the full covariance matrix — the pairwise correlations and volatilities of every holding. Non-linear portfolios containing options are poorly served by simple parametric VaR, which captures only first-order (delta) sensitivity to price changes.

Advantages: Fast, closed-form, easy to implement for linear portfolios.

Disadvantages: Assumes normality — underestimates risk when returns have fat tails or negative skewness.

Historical Simulation VaR

Historical simulation avoids distributional assumptions entirely. Instead, it uses actual past returns to build the loss distribution:

1. Collect N historical portfolio returns (e.g., 500 daily returns)
2. Sort returns from worst to best
3. Find the return at the desired percentile cutoff

For 95% VaR with 1,000 observations, the VaR is the 50th worst return (the 5th percentile). For 99% VaR, it is the 10th worst return.

Advantages: Captures fat tails, skewness, and non-normal features actually present in historical data. No model assumptions required.

Disadvantages: Limited to the historical window observed — cannot capture scenarios that haven’t occurred yet. Sensitive to sample period selection. For related historical risk analysis, see maximum drawdown.

Monte Carlo VaR

Monte Carlo VaR generates thousands of simulated portfolio return scenarios to construct a loss distribution. It is the most flexible of the three methods:

1. Define return distributions and correlations for all portfolio holdings
2. Generate N random scenarios (typically 10,000+)
3. Value the portfolio under each scenario
4. Sort the resulting profit-and-loss (P&L) outcomes
5. Find the VaR at the desired percentile

Advantages: Handles non-normal distributions, complex instruments (options, structured products), and non-linear payoffs. Can incorporate forward-looking assumptions.

Disadvantages: Computationally intensive. Results depend entirely on the input assumptions — “garbage in, garbage out.” For more on Monte Carlo simulation in finance, see the dedicated article.

Value at Risk Example

Consider a $1,000,000 U.S. equity portfolio with an annual volatility of 15% (approximately the long-term volatility of the S&P 500).

The three methods can produce different VaR estimates for the same portfolio, because each relies on different assumptions:

Historical and Monte Carlo results are approximate ranges. The differences illustrate why many institutions calculate VaR using multiple methods and compare results.

VaR Backtesting

A VaR model is only useful if it accurately predicts the frequency of losses exceeding the VaR threshold. Backtesting validates VaR models by counting the number of days actual portfolio losses exceeded the predicted VaR — these are called exceptions or breaches.

For a 99% 1-day VaR model evaluated over 250 trading days (approximately one year), you would expect roughly 2.5 exceptions. If the number of exceptions is significantly higher, the model underestimates risk. If significantly lower, the model may be too conservative, leading to capital inefficiency.

Value at Risk vs Expected Shortfall (CVaR)

VaR and Expected Shortfall (also called Conditional VaR or CVaR) are complementary risk measures. VaR identifies the loss threshold; CVaR tells you what happens beyond it.

The critical distinction is tail blindness. Two portfolios can have identical VaR but vastly different tail risk profiles. Portfolio A might lose slightly more than VaR on bad days, while Portfolio B — perhaps concentrated in illiquid or leveraged positions — could suffer catastrophic losses. CVaR distinguishes between them by averaging all losses beyond the VaR threshold. For a deeper treatment of CVaR methodology, see the Expected Shortfall article.

VaR is best understood alongside other risk metrics that capture different dimensions of portfolio risk:

Common Mistakes

These are the most frequent errors practitioners and students make when working with VaR:

1. Treating VaR as worst-case loss — VaR is not the maximum possible loss. It is a threshold: losses beyond VaR are expected to occur (1-α)% of the time, and those losses can be far larger. In the 2008 crisis, many portfolios experienced losses multiples of their reported VaR.

2. Mixing confidence levels when comparing — Comparing a 95% VaR to a 99% VaR across portfolios is meaningless without adjustment. A fund reporting 95% VaR of $1M may have the same or more risk than a fund reporting 99% VaR of $1.5M. Always use the same confidence level and time horizon when comparing.

3. Ignoring tail risk — VaR says nothing about the magnitude of losses beyond the threshold. A portfolio with 99% VaR of $100,000 could lose $110,000 or $500,000 in a tail event — VaR does not distinguish between them. Use Expected Shortfall (CVaR) to assess tail severity.

4. Assuming returns are normally distributed — Parametric VaR assumes normality, but equity returns exhibit fat tails and negative skewness. At the 99% confidence level, the normal distribution predicts a 2.33-sigma event, but actual equity markets produce losses of this magnitude more frequently. This causes parametric VaR to systematically underestimate true risk.

5. Using two-tailed z-scores instead of one-tailed — VaR measures the left tail only (losses). Using a two-tailed z-score — for example, 1.96 for 95% confidence instead of the correct 1.645 — overstates VaR by applying a more extreme quantile than intended. Always use one-tailed critical values.

6. Mismatching horizon and frequency — Plugging annual volatility directly into a 1-day VaR formula without converting (dividing by √252) dramatically overstates risk. A 15% annual σ is not a 15% daily σ — the daily figure is approximately 0.945%. Always ensure the volatility time unit matches the VaR horizon.

7. Using too short a data window — Historical simulation VaR based on only 1–2 years of calm market data will systematically underestimate risk. If the data window excludes a recession or market crisis, the VaR model has never “seen” a stress scenario. Use a data window that includes at least one stress regime — often 3 to 5+ years in practice — for reliable estimates.

Limitations of Value at Risk

1. Tail blindness — VaR provides no information about the shape or magnitude of the loss distribution beyond the confidence threshold. Two portfolios with the same VaR can have vastly different tail exposures. This fundamental limitation motivated the development of Expected Shortfall (CVaR) as a tail-aware alternative.

2. Not sub-additive — A coherent risk measure should satisfy: Risk(A+B) ≤ Risk(A) + Risk(B), meaning diversification should never appear to increase risk. VaR violates this property — in certain scenarios, the VaR of a combined portfolio can exceed the sum of the individual VaRs. This is theoretically unsatisfying and practically problematic for firms that aggregate risk across business units.

3. Assumption sensitivity — Parametric VaR depends on the normality assumption. Historical VaR depends on which sample period is chosen. Monte Carlo VaR depends on the chosen model and its calibration. All three methods can produce materially different VaR estimates for the same portfolio, and there is no definitive way to determine which is “correct.”

4. Procyclicality — VaR-based capital requirements fall during calm periods (when measured volatility is low) and spike during crises (when volatility surges). This procyclical behavior can amplify market stress: just when firms most need to hold positions, rising VaR forces them to reduce risk — potentially accelerating the sell-off.

5. False precision — A single VaR number can create a false sense of certainty. Reporting “99% VaR is $22,000” suggests more precision than the underlying model warrants. Responsible risk management complements VaR with stress testing, scenario analysis, and Expected Shortfall to capture what VaR misses.