VaR vs. CVaR (Expected Shortfall): Which Measures Tail Risk Better?

Value-at-Risk tells you the worst reasonable loss at a confidence level. But it doesn't tell you what happens when things go worse. A portfolio's 99% VaR might be $10,000, meaning there's a 1% chance of losing more than $10,000. But on those bad 1% days, do you typically lose $10,500 or $30,000? VaR is silent on this crucial question.

Conditional Value-at-Risk (CVaR), also called Expected Shortfall (ES), answers that gap. It measures the average loss on days when you exceed your VaR threshold. If your 99% VaR is $10,000 but your 99% CVaR is $15,000, you now know that on bad days (the 1% tail), you expect to lose $15,000 on average. This extra information shapes position sizing and stress-testing decisions. This article explains CVaR, compares it to VaR, and shows when each is appropriate.

Quick definition: Conditional Value-at-Risk (CVaR), or Expected Shortfall, is the average loss that occurs when losses exceed your VaR threshold, measuring the size of tail-end losses.

Key takeaways

• CVaR measures average loss beyond the VaR threshold, addressing VaR's blindness to tail size
• A portfolio can have the same VaR but very different CVaR depending on the tail shape
• CVaR is more conservative and better for stress scenarios and capital allocation
• Regulatory trend is toward CVaR; Basel III increasingly emphasizes it
• VaR remains useful for position sizing and intraday monitoring despite CVaR's advantages

The Core Problem VaR Leaves Unsolved

Imagine two portfolios, each with a $100,000 value and a 99% confidence VaR of $5,000. They appear equally risky under VaR. But examine the 1% worst case more closely:

Portfolio A: On the 1% bad days, losses average $6,200 Portfolio B: On the 1% bad days, losses average $12,400

These portfolios have identical VaR but vastly different tail risk. Portfolio B is far riskier because when things go wrong, they go very wrong. VaR's failure to distinguish this difference is a critical limitation.

This scenario isn't theoretical. A portfolio with many small positions and uncorrelated assets might have a small tail (like Portfolio A). A portfolio with a concentrated short option position has a large tail (like Portfolio B). Both could have the same VaR, but a risk manager would size them very differently.

CVaR fills this gap. CVaR for Portfolio A would be $6,200 (close to VaR). CVaR for Portfolio B would be $12,400 (far larger than VaR). Now the risk is comparable to the actual tail behavior.

Calculating CVaR: The Method

CVaR is straightforward to calculate once you have your loss distribution.

In historical VaR terms: Rank your historical returns from worst to best. Identify the VaR threshold (e.g., the 3rd worst day for 99% confidence). Take all returns that are worse than (or equal to) that threshold and average them.

Example: 252 trading days, ranked from worst to best:

Rank 1: -4.5% return
Rank 2: -3.9%
Rank 3: -3.2% ← This is your 99% VaR threshold
Rank 4: -2.1%
Rank 5: -1.8%
...
Rank 252: +4.3%

Your 99% VaR is -3.2%. For CVaR, take the average of all returns at or worse than -3.2%:

99% CVaR = (Rank 1 + Rank 2 + Rank 3) / 3
          = (-4.5 + -3.9 + -3.2) / 3
          = -11.6 / 3
          = -3.87%

On a $100,000 portfolio:

• 99% VaR: $3,200
• 99% CVaR: $3,870

The $670 difference might seem small, but it scales. On a $10 million portfolio, it's $67,000—material when sizing positions or setting capital reserves.

Parametric CVaR: Using the Normal Distribution

If returns are assumed normal (as in parametric VaR), CVaR can be calculated analytically:

CVaR = μ - σ × φ(Z) / (1 - confidence%)

Where:

• μ is expected return
• σ is standard deviation
• φ(Z) is the normal probability density function at the Z-score
• confidence% is your confidence level (e.g., 0.99)

For a portfolio with 0.05% daily expected return and 1.5% daily volatility, at 99% confidence (Z = 2.326):

Normal PDF at Z=2.326: φ(2.326) ≈ 0.0273

CVaR = 0.0005 - 0.015 × 0.0273 / (1 - 0.99)
     = 0.0005 - 0.015 × 0.0273 / 0.01
     = 0.0005 - 0.04095
     = -0.04045 = -4.045%

On a $100,000 portfolio:

• Parametric 99% VaR: $100,000 × 2.326 × 0.015 = $3,489
• Parametric 99% CVaR: $100,000 × 4.045% = $4,045

CVaR is 16% larger than VaR, a proportional difference that depends on the normal distribution shape.

Monte Carlo CVaR: Direct from Simulation

In Monte Carlo, calculate CVaR from the same simulations used for VaR. After running 10,000 simulations:

1. Rank the 10,000 simulated returns from worst to best.
2. Identify the 1% worst outcomes (roughly the 100th worst simulation out of 10,000).
3. Take the average return of those 100 worst simulations.

This average is your Monte Carlo CVaR. The advantage of Monte Carlo is that CVaR emerges naturally from the tail of the distribution without parametric assumptions. If the tail is fat (more extreme outcomes), CVaR will capture that directly.

Real-World Example: A Stock Portfolio vs. An Options Portfolio

Compare two $250,000 portfolios:

Portfolio A (Stocks): 80% SPY, 20% QQQ. Returns are reasonably normal.

Historical analysis of 252 days:

• 95% VaR: -$4,200

• 95% CVaR: -$5,100 (about 21% larger)

• 99% VaR: -$6,800

• 99% CVaR: -$8,400 (about 24% larger)

Portfolio B (Options-enhanced): 70% SPY, 20% QQQ, and short 50 call option spreads (bet on low volatility). This portfolio has a large negative skew; when it loses, it loses catastrophically.

Historical analysis of the same 252 days:

• 95% VaR: -$5,600

• 95% CVaR: -$9,200 (about 64% larger)

• 99% VaR: -$9,100

• 99% CVaR: -$18,500 (about 103% larger)

Portfolio B's VaR looks only moderately worse than Portfolio A. But the CVaR reveals that on bad days, Portfolio B loses far more. A risk manager would size Portfolio B's positions much smaller than Portfolio A's despite similar VaR, because the tail risk is so much worse.

The Advantages of CVaR Over VaR

Coherent risk measure: CVaR is mathematically "coherent," meaning it satisfies properties that VaR violates. For example, diversification always reduces CVaR (adding an uncorrelated asset can't increase CVaR), which is not always true for VaR. This makes CVaR more theoretically sound.

Addresses tail risk: VaR is blind to tail size; CVaR isn't. This is crucial for stress-testing and capital allocation.

Stress-friendly: When you ask "what's the worst realistic loss?" VaR answers with a threshold, CVaR answers with the average of bad outcomes, which is more actionable for position sizing.

Regulatory preference: After the 2008 crisis, regulators increasingly prefer CVaR. The Basel III accord explicitly mentions Expected Shortfall. The SEC guidance on hedge fund risk reporting is shifting toward CVaR alongside VaR.

Incentive alignment: VaR incentivizes "probability management"—reducing the chance of losses. CVaR incentivizes "tail management"—reducing the size of losses when they occur. This is more directly aligned with risk governance objectives.

The Limitations of CVaR

Computational complexity: CVaR requires more data or simulations to estimate reliably. Historical CVaR needs you to have several observations beyond the VaR threshold. With 252 days and 99% confidence (roughly 2-3 observations), CVaR estimation is noisy.

Less intuitive: VaR is easier to explain. "There's a 95% chance we won't lose more than $5,000" is clearer than "our expected loss on bad days is $7,200." Most traders and executives still think in terms of VaR first.

Parameter sensitivity: CVaR is more sensitive to tail estimation errors. A single incorrect extreme observation can skew CVaR more than it would skew VaR.

Not a complete tail measure: Even CVaR doesn't measure the absolute worst case. If your 99% CVaR is $10,000 but there's a 0.01% chance of a $50,000 loss, CVaR won't flag this. This is why stress-testing complements both VaR and CVaR.

Comparing the Three Calculation Methods

Aspect	Historical	Parametric	Monte Carlo
CVaR Calculation	Average tail returns	Formula	Average tail simulations
Data requirement	252+ days	252 days	10,000+ sims
Assumes normality	No	Yes	Depends on model
Handles fat tails	Yes	No	Yes if sims include them
Works for options	Yes	Poor	Yes
Speed	Fast	Very fast	Moderate

When to Use VaR vs. CVaR

Use VaR for:

• Daily position sizing and trading limits
• Routine intraday monitoring
• Quick risk dashboards
• Regulatory compliance (when regulators explicitly require VaR)
• Communicating risk to non-technical stakeholders

Use CVaR for:

• Stress scenario analysis ("What's a realistic worst day?")
• Long-term capital allocation
• Options and derivatives portfolios
• Portfolio comparison (when VaR is similar but tail risk differs)
• Hedge fund and bank regulatory reporting
• Performance attribution (measuring risk-adjusted returns)

Use both together for:

• Comprehensive risk governance
• Identifying portfolios with hidden tail risk
• Backtesting and model validation

Real Example: Portfolio Risk Governance

A fund manager oversees three sub-portfolios. She calculates VaR and CVaR for each at 99% confidence, 1-day horizon:

Portfolio	Value	VaR	CVaR
Equity Core	$10M	-$185K	-$240K
Emerging Markets	$5M	-$95K	-$160K
Option Strategies	$3M	-$80K	-$210K
Total	$18M	-$320K	-$520K

The Option Strategies portfolio appears least risky by VaR ($80K on $3M = 2.67%). But by CVaR ($210K on $3M = 7%), it's far riskier. The manager reduces the option portfolio to $1.5M, cutting its position size in half. Now:

Portfolio	Value	VaR	CVaR
Equity Core	$10M	-$185K	-$240K
Emerging Markets	$5M	-$95K	-$160K
Option Strategies	$1.5M	-$40K	-$105K
Total	$16.5M	-$255K	-$400K

The reduction in CVaR (from $520K to $400K) is proportionally larger than the reduction in VaR, protecting against tail events.

Common Mistakes with CVaR

Using CVaR alone without VaR context: CVaR can obscure probability. A portfolio might have very low CVaR if bad days are rare but severe. You need both metrics.

Underestimating CVaR estimation error: With limited tail observations, CVaR estimates are noisy. A 99% CVaR based on only 3 observations (252 × 0.01) is unreliable. Use longer lookbacks, simulations, or multiple estimation methods.

Confusing CVaR with worst case: CVaR is the average of bad days, not the absolute worst day. If your worst historical loss was -12%, but your CVaR is only -8%, that's correct; it's the average, not the maximum.

Applying CVaR to illiquid portfolios: CVaR assumes you measure historical returns accurately. For illiquid assets with stale pricing or gaps, the historical distribution is misleading.

FAQ

How do I estimate CVaR if I have few tail observations?

Increase your lookback period (use 500-1000 days if available) or use Monte Carlo simulation (which generates many tail outcomes). Historical CVaR with fewer than 5 observations in the tail is unreliable.

Is CVaR always larger than VaR?

Yes, by mathematical definition. CVaR averages all outcomes at or worse than the VaR threshold, so it's always equal to or worse than the threshold. The gap varies: it can be small (thin tail) or large (fat tail).

How does CVaR scale with time?

Like VaR, it scales with the square root of time for parametric models. A 1-day 99% CVaR of $5,000 becomes approximately $5,000 × sqrt(10) = $15,811 for a 10-day horizon, assuming constant volatility and correlation.

Can I use CVaR for portfolio optimization?

Yes, increasingly. "CVaR-minimizing portfolios" are an alternative to traditional mean-variance optimization. They tend to be more conservative and robust to tail risk than minimum-variance portfolios.

How do I backtest CVaR?

Track your CVaR estimate each day. On days when actual losses exceed VaR, record the excess. Over time, the average excess should approximately equal your CVaR estimate minus your VaR estimate. If actual tail losses are much larger, your tail estimation is optimistic.

Should I report both VaR and CVaR to stakeholders?

Yes, if possible. Report VaR as the primary metric (it's what people understand), then footnote CVaR with an explanation: "On the 1% of days when losses exceed our $5K VaR, we typically lose about $8K on average (CVaR)."

What if my CVaR is negative but less severe than VaR suggests?

This shouldn't happen; CVaR should always be worse than VaR (larger loss magnitude). If you see the opposite, you have a calculation error. Check that you're averaging the right tail outcomes.

Is there a CVaR equivalent for upside (gains)?

Yes, sometimes called "Conditional Value-at-Gain" or "Expected Shortfall for gains." It's the average gain on your best days. It's less commonly used but can be useful for upside-focused investors.

Do regulators require CVaR reporting?

Not universally yet. Basel III references Expected Shortfall, but most banks still report VaR. The SEC doesn't mandate CVaR for hedge funds, though the trend is toward it. Check your regulator's current guidance.

How do I explain CVaR to a CEO who thinks VaR is already too conservative?

Frame it as tail-risk specificity: "VaR tells us the threshold of a bad day. CVaR tells us the actual average loss when we hit that threshold. If we're right about our VaR but wrong about the tail shape, we could lose far more than planned on crisis days. CVaR helps us size positions so we can survive even our worst bad days."

Related concepts

• What Is Value-at-Risk?
• Parametric VaR: The Variance-Covariance Method
• Historical VaR: Using Past Returns
• What Is a Black Swan?

Summary

Conditional Value-at-Risk (Expected Shortfall) measures the average loss beyond your VaR threshold, addressing VaR's critical blind spot: it doesn't tell you how bad tail losses actually are. A portfolio can have the same VaR as another but very different CVaR if its tail is fatter. This difference is crucial for position sizing, stress testing, and capital allocation. CVaR is mathematically more coherent than VaR and increasingly required by regulators. However, it's harder to estimate from limited data and less intuitive for daily communication. Professional risk teams use both metrics: VaR for routine position sizing and intraday monitoring, CVaR for stress scenarios and long-term risk governance. Together, they provide a more complete picture of portfolio risk—both the probability of loss and the severity of that loss when it occurs.

Next

→ 1-Day vs. 10-Day VaR: Which to Use?