Value at Risk (VaR) answers a single question: "What is the most I could lose, over a given time period, in all but the worst-case scenarios?" It produces a dollar figure or a percentage that summarizes a portfolio's downside risk into one number. For example, a one-day 95% VaR of $100,000 means that on 95 out of 100 trading days, the portfolio is expected to lose less than $100,000.

VaR became the standard risk measure in banking and investment management during the 1990s, driven largely by JPMorgan's RiskMetrics framework. It is now embedded in financial regulation: the Basel Accords require banks to calculate VaR to determine how much capital they must hold against trading losses. Despite its widespread adoption, VaR has significant limitations that became painfully clear during the 2008 financial crisis, when actual losses far exceeded VaR estimates at many institutions.

Conceptual Framework

VaR is a quantile measure of the loss distribution. At its core, it asks: "If I rank all possible outcomes from worst to best, what loss level separates the worst 5% (or 1%) from the rest?" This framing makes VaR intuitive for non-technical audiences, which is one reason it became so popular. A board of directors or a risk committee can understand "We could lose up to $50 million on a bad day" more readily than a covariance matrix or a volatility estimate.

The calculation requires three inputs: a confidence level (typically 95% or 99%), a time horizon (one day, ten days, or one month), and a method for estimating the distribution of portfolio returns. The confidence level determines how far into the tail of the distribution the measure reaches. A 99% VaR is more conservative than a 95% VaR because it captures rarer, more extreme events.

Core Assumptions

All VaR calculations share certain assumptions, though the specific assumptions depend on the estimation method used:

• The past informs the future: Every VaR method uses historical data, either directly (historical simulation) or to estimate parameters (parametric VaR) or to calibrate a model (Monte Carlo VaR). This means VaR is backward-looking by construction. If the future behaves differently from the past, VaR will be wrong.
• Portfolio composition is static: Standard VaR calculations assume the portfolio does not change during the holding period. In reality, traders adjust positions in response to market movements. This assumption matters more for longer horizons (ten-day VaR) than for overnight calculations.
• Markets remain liquid: VaR assumes that positions can be valued at current market prices. During a crisis, bid-ask spreads widen dramatically and some markets become illiquid. The actual cost of exiting positions under stress can far exceed the loss implied by VaR, which uses normal market pricing.
• VaR says nothing about the size of tail losses: A 95% VaR tells you the boundary of the worst 5%, but it says nothing about what happens within that worst 5%. The portfolio could lose $100,000 or $500,000 in the tail, and VaR treats both scenarios identically. This is the most fundamental criticism of VaR as a risk measure.

Estimation Methods

Three primary methods are used to calculate VaR. Each makes different tradeoffs between simplicity, accuracy, and computational cost.

Parametric VaR (Variance-Covariance)

The parametric method assumes that portfolio returns follow a bell curve (normal distribution). Under this assumption, VaR can be calculated with a simple formula: multiply the portfolio's standard deviation by a z-score corresponding to the confidence level, then multiply by the portfolio value.

For a 95% confidence level, the z-score is 1.65; for 99%, it is 2.33. A $10 million portfolio with 1% daily volatility (standard deviation) would have a one-day 95% parametric VaR of $10,000,000 x 0.01 x 1.65 = $165,000.

The advantage is speed: once the covariance matrix is estimated, the calculation is nearly instantaneous. The disadvantage is the normality assumption. Financial returns are not normally distributed; they have fatter tails (more extreme events) than a bell curve predicts. Parametric VaR systematically underestimates the risk of large losses.

Historical Simulation

Historical simulation takes the current portfolio and applies actual past returns to it, day by day. If using 500 days of history, it produces 500 hypothetical daily profit-and-loss outcomes. VaR is simply the loss at the appropriate percentile of this distribution. For 95% VaR with 500 observations, the VaR is the 25th-worst loss (the 5th percentile).

This method makes no assumption about the shape of the return distribution, which is a significant advantage over the parametric approach. It automatically captures the fat tails, skewness, and non-linearities present in historical data. The main weakness is that it is entirely dependent on the historical window chosen. If the window does not include a market stress event, the VaR estimate will be artificially low. Additionally, all past observations are weighted equally, even though recent market behavior may be more relevant than behavior from years ago.

Monte Carlo VaR

Monte Carlo VaR uses random simulation to generate thousands of hypothetical return scenarios. A statistical model describes how each risk factor (interest rates, stock prices, exchange rates) behaves and how they relate to each other. The simulation draws random values from this model, revalues the portfolio under each scenario, and produces a distribution of portfolio outcomes. VaR is then read from the tail of this simulated distribution.

Monte Carlo is the most flexible method. It can handle complex portfolios with options and other non-linear instruments, and it can incorporate fat-tailed distributions, time-varying volatility, and changing correlations. The cost is computational intensity: a large portfolio may require millions of simulations, and the results are only as good as the statistical model used to generate the scenarios.

Risk Architecture

VaR is a risk measurement tool, but it introduces its own set of risks when used as the primary risk management metric. The 2008 financial crisis exposed several of these weaknesses.

Model Risk

The most critical model risk is tail blindness. VaR tells you the boundary of normal losses but nothing about what happens beyond that boundary. Two portfolios can have identical VaR numbers but vastly different tail risks. Portfolio A might have maximum losses only slightly worse than its VaR, while Portfolio B might face catastrophic losses in the tail. VaR treats them as equally risky.

Expected Shortfall (also called Conditional VaR or CVaR) addresses this limitation by measuring the average loss in the tail beyond the VaR threshold. If the 95% VaR is $100,000, Expected Shortfall asks: "When losses exceed $100,000, what is the average loss?" This provides information about the severity of tail events, not just their boundary. The Basel III framework (FRTB) now requires banks to use Expected Shortfall instead of VaR for market risk capital calculations.

Known Limitations

Practical Considerations

Choosing the Confidence Level

The choice between 95% and 99% confidence levels involves a tradeoff between sensitivity and stability. A 95% VaR is exceeded roughly once per month (one trading day in 20), providing enough breach events to validate the model through backtesting. A 99% VaR is exceeded roughly 2-3 times per year, making model validation more difficult because breaches are rare.

Regulatory VaR uses a 99% confidence level with a ten-day holding period. Internal risk management often uses 95% with a one-day horizon because the more frequent breaches allow the risk team to evaluate model accuracy more quickly.

Time Horizon Scaling

Converting VaR from one time horizon to another is common practice. The simplest approach multiplies the one-day VaR by the square root of the number of days. A one-day VaR of $100,000 becomes a ten-day VaR of $100,000 x √10 ≈ $316,000. This "square root of time" rule assumes that returns are independent from one day to the next and that volatility is constant. Both assumptions are violated in practice. Volatility clustering (big moves followed by big moves) and serial correlation in returns make the square root rule approximate at best.

Backtesting VaR Models

VaR models must be validated by comparing predicted VaR to actual outcomes. This process, called backtesting, counts how often the actual portfolio loss exceeded the VaR estimate. For a 99% VaR calculated over 250 trading days, the expected number of breaches is 2.5 per year. Significantly more breaches suggest the model underestimates risk; significantly fewer suggest overestimation.

The Basel framework uses a traffic-light system: zero to four breaches in 250 days is "green" (acceptable), five to nine is "yellow" (requiring investigation), and ten or more is "red" (requiring a capital multiplier). This backtesting requirement creates an incentive for banks to use conservative VaR models, but overly conservative models tie up capital unnecessarily.

Expected Shortfall as an Alternative

Expected Shortfall (ES), sometimes called Conditional Value at Risk (CVaR), measures the average loss in the tail beyond VaR. If the 95% VaR is $100,000, the 95% ES might be $150,000, meaning that when losses exceed $100,000, they average $150,000. ES addresses VaR's biggest weakness by providing information about tail severity. It is also mathematically "coherent," meaning it consistently shows that diversification reduces risk, a property VaR lacks under certain conditions.

Related Models

Further Reading

• Jorion, P. (2007). Value at Risk: The New Benchmark for Managing Financial Risk (3rd ed.). McGraw-Hill.
• Dowd, K. (2005). Measuring Market Risk (2nd ed.). Wiley.
• Artzner, P., Delbaen, F., Eber, J.M. and Heath, D. (1999). "Coherent Measures of Risk." Mathematical Finance, 9(3), 203–228.
• Basel Committee on Banking Supervision (2019). "Minimum Capital Requirements for Market Risk." Bank for International Settlements.
• McNeil, A.J., Frey, R. and Embrechts, P. (2015). Quantitative Risk Management: Concepts, Techniques and Tools (2nd ed.). Princeton University Press.
• Glasserman, P. (2003). Monte Carlo Methods in Financial Engineering. Springer (Stochastic Modelling and Applied Probability, Vol. 53).