Value at Risk Explained: How Banks Measure Tail Risk

Value at Risk (VaR) is finance’s most-used and most-criticized risk number. It compresses the question “how much could this portfolio lose tomorrow?” into a single dollar figure tied to a confidence level. Every major bank reports it, every Basel and SEC rule references it, and every textbook teaches it – which makes its blind spot (it cannot tell you anything about losses beyond the threshold) one of the most important things to understand about modern risk management.

TL;DR: The 1-day 99% VaR of a $10 million portfolio is the loss level that should be exceeded only about 1 trading day in 100. If the number is $230,000, you are saying: “In a normal day, our worst credible loss is around $230k; on the 1% of days that are not normal, we have no idea how bad it gets.” That last clause is why regulators have shifted toward Expected Shortfall, which averages the size of losses beyond the VaR cutoff.

The formal definition

For a portfolio with random loss L over a chosen horizon, the Value at Risk at confidence level α (for example 95% or 99%) is the smallest threshold the loss is unlikely to exceed:

VaRα(L) = inf { x : P(L > x) ≤ 1 − α }

Plain English: pick a horizon (1 day, 10 days), pick a confidence level (95%, 99%, 99.5%), and find the loss number such that the probability of doing worse is only (1 − α). The three pieces of any VaR quote are therefore horizon, confidence, and the resulting dollar (or percentage) loss. A VaR with no horizon or no confidence attached is meaningless.

Three things VaR does not tell you, no matter how it is computed:

  • How bad the bad days are. VaR is the threshold of the tail, not the average of the tail.
  • Anything about the diversification. VaR is not subadditive, which means the VaR of a combined portfolio can be larger than the sum of the parts’ VaRs. That violates a basic property risk theorists want from a coherent risk measure.
  • Anything about regimes. A VaR calibrated to calm markets will systematically understate risk going into a crisis.

The three ways to compute VaR

Banks and funds use one (or all three) of these methods. The BIS’s overview of the regulatory framework groups VaR techniques into the same three families, and the McKinsey-Risk industry survey cited in academic reviews of VaR practice has long shown historical simulation dominating large banks’ usage with Monte Carlo a distant second.

Method How it works Strengths Weaknesses
Historical simulation Re-price today’s portfolio under the last N days of observed market moves. Sort the resulting P&Ls; the (1 − α) quantile is the VaR. No distributional assumptions. Captures real, fat-tailed history. Easy to explain to non-quants. Past 1-3 years rarely contain the worst-case move. Sensitive to the window chosen.
Variance-covariance (parametric) Assume returns are normally distributed. VaR = zα × σ × portfolio value, where zα is the standard-normal quantile. Closed-form and fast. Works as a sanity check. Normality assumption fails in real markets – actual return distributions are fat-tailed and skewed. Systematically understates extreme losses.
Monte Carlo simulation Draw thousands of synthetic market-move paths from a calibrated multivariate distribution; re-price the portfolio under each; take the quantile of simulated losses. Handles non-linear positions (options, convex bonds) and arbitrary distributions. The only practical option for complex derivative books. Expensive. Result is only as good as the assumed joint distribution. Garbage in, garbage out.
Sources: BIS, Minimum capital requirements for market risk (Jan 2019). Wikipedia, Value at risk. McKinsey’s market-risk survey cited in industry reviews of VaR practice.

The standard analogy: historical simulation says “yesterday’s worst day is roughly our worst day.” Parametric VaR says “if returns were a clean bell curve, here is the cutoff.” Monte Carlo says “let us write down a richer model and roll the dice ten million times.” Each is a different bet on which lie about the world is least harmful.

A worked example with real numbers

Suppose you hold a $10 million S&P 500 ETF position and want a 1-day 99% VaR. Use the parametric method as the textbook starting point. The S&P 500’s daily return standard deviation in a typical year sits near 1.0% (it spent most of 2024-2026 between 0.7% and 1.2%; in crisis weeks it spiked above 4%). The standard normal quantile at 99% is z0.99 ≈ 2.326.

1-day 99% VaR ≈ 2.326 × 0.010 × $10,000,000 = $232,600

Drop the confidence to 95% and z falls to 1.645, so the VaR drops to about $164,500. Extend the horizon to 10 trading days under the square-root-of-time rule used in Basel’s traffic-light backtest, and the 99% number grows by √10 to roughly $735,500. Already you can see two of the dangers of parametric VaR: the number is highly sensitive to the σ you plug in (real volatility moves a lot), and the square-root-of-time scaling assumes returns are independent across days, which they are not when markets trend or crash.

VaR as a quantile of the loss distribution A bell-curve-shaped loss distribution. The 95% and 99% VaR points are marked as vertical lines on the right tail; the area beyond the 99% line is shaded as the tail VaR does not measure.

1-day loss distribution: where VaR sits

density loss → gain ←

95% VaR ($164,500)

99% VaR ($232,600)

tail VaR ignores

$10M S&P 500 position, σ = 1.0%/day, normal distribution assumed

Schematic. VaR is the quantile cutoff; Expected Shortfall is the conditional mean of the shaded tail.

Real history overflows the parametric numbers

The parametric VaR above implies the S&P 500 should drop more than 2.33 standard deviations only about 1 day in 100. Reality has many days well outside that boundary:

Date Event S&P 500 1-day return Std devs from zero (σ = 1.0%)
Oct 19, 1987 Black Monday −20.5% ~20σ
Oct 15, 2008 Great Financial Crisis −9.0% ~9σ
Mar 16, 2020 COVID shock −12.0% ~12σ
Aug 8, 2011 US debt downgrade −6.7% ~7σ
Sources: closing prices via FRED – S&P 500 (daily) and Federal Reserve History – Stock Market Crash of 1987. The σ column uses the textbook 1.0% daily benchmark and is intentionally rough.

A 20-sigma event under a normal distribution should not occur once in the age of the universe. The S&P 500 has produced several of them inside a single lifetime. This is the well-known fat-tail problem, and it is why a parametric 99% VaR can pass 99 days out of 100 and still be wrong about the loss that actually matters.

Why VaR cannot tell you about the tail

Set aside method choice for a moment. VaR has a structural blind spot baked into its definition. By construction it is the threshold of the tail, not the depth. Two portfolios can have identical 99% VaRs and very different consequences if the worst 1% of days actually arrives.

VaR versus Expected Shortfall on the same tail Two histogram-style return distributions side by side. Both have the same 99% VaR cutoff. The left one has a clean tail beyond the cutoff; the right one has a fat tail with much larger losses beyond the cutoff. Expected Shortfall captures the difference; VaR does not.

Same 99% VaR, very different tails

Portfolio A (thin tail) 99% VaR small tail

Portfolio B (fat tail) 99% VaR fat tail

VaR sees the line. Expected Shortfall sees the area to the right of the line.

Schematic. Expected Shortfall (ES, also called CVaR or Conditional VaR) is the conditional expectation E[L | L > VaR]. It answers “if we end up in the worst 1% of days, how bad is the average bad day?”

The famous David Einhorn quote cited in the Wikipedia article on VaR puts it bluntly: VaR “focused on the manageable risks near the center of the distribution and ignored the tails.” That is not a complaint about a particular calculation method – it is a complaint about the metric itself. Even a perfect 99% VaR by construction says nothing about the size of the worst 1%.

How regulators responded: from VaR to Expected Shortfall

The 2008 crisis embarrassed the parametric, calm-market-calibrated VaR numbers banks had been reporting. The Basel Committee’s 2019 Minimum Capital Requirements for Market Risk (the final standard from the Fundamental Review of the Trading Book, or FRTB) replaced the old VaR-based internal-models approach with an Expected-Shortfall-based one and added a stressed calibration so capital charges no longer collapse during a quiet year. The Fed and other supervisors followed with national implementations. The headline change for risk officers: the IMA capital charge is now anchored to the average of the bad tail, not its threshold.

Common mistakes when reading or quoting VaR

  • Quoting it without horizon or confidence. “Our VaR is $5M” is meaningless. “Our 1-day 99% VaR is $5M” is a claim.
  • Treating the quantile as the worst case. A 99% VaR is exceeded on about 2-3 trading days a year. That is not the worst case; it is the merely-bad case.
  • Trusting parametric numbers for option-heavy books. Options have non-linear, asymmetric payoffs. A normal-distribution VaR badly misrepresents the loss distribution of even a simple short-put position.
  • Forgetting backtesting. Basel’s traffic-light system (green / yellow / red zones) is a count of exceptions per 250 trading days. A green model can still understate severity – which is partly why FRTB moved past it.
  • Mistaking VaR for diversification math. VaR is not subadditive. Combining two desks can produce a VaR larger than the sum of the individual VaRs.

Where this leaves a working investor

VaR remains the dominant industry shorthand because it is simple to quote and easy to map to a capital number. Treat it as a useful summary in normal conditions and a misleading one in tail conditions. If the only number on a fund’s risk page is a VaR, the next questions are: which method, which window, which confidence level, what was the most recent backtest, and what does the Expected Shortfall look like alongside it? Pair VaR with at least one tail-aware measure and one drawdown-aware measure, and the resulting picture is far more honest than the headline number alone.

Related concepts and what to learn next

  • Expected Shortfall / CVaR / Conditional VaR – the average loss conditional on being in the tail. Coherent (subadditive). The Basel FRTB metric.
  • Stress testing – scenario-based losses (1987, 2008, COVID, hypothetical rates spike). Complements statistical VaR with named, plausible disasters. The Fed’s CCAR/DFAST framework for large banks is the most visible US example.
  • Backtesting – count actual exceptions versus expected. Required by regulators; useful in-house for catching model drift.
  • Maximum drawdown – the realized worst peak-to-trough loss. See the explainer. A complementary, path-dependent risk number.
  • Sharpe ratio – return per unit of volatility. See the explainer. What VaR-aware investors pair with a tail measure.

The short version

Value at Risk turns the question “how much could we lose?” into a one-number answer at a chosen horizon and confidence. Compute it three ways – historical, parametric, or Monte Carlo – knowing each is a different bet on the world. Quote it with horizon and confidence or do not quote it. And remember the structural blind spot: VaR is the threshold of the tail, not its depth. The Basel response to 2008 was to move regulatory capital from VaR to Expected Shortfall, and that move tells you everything about how to use VaR in practice – as a useful summary, never as a guarantee.

Sources & further reading

Disclosure: This article was produced with AI assistance and reviewed before publication. It is for informational purposes only and is not investment advice.

Leave a Comment