Conditional Value at Risk (CVaR), also known as Expected Shortfall or Expected Tail Loss, measures the average loss a portfolio would experience specifically within its worst-case tail of outcomes — the losses that occur beyond the standard Value at Risk (VaR) threshold. Where VaR answers “what loss level will I exceed only X% of the time,” CVaR answers the more consequential follow-up question: “given that I am in that worst X% of outcomes, how bad should I expect it to be, on average?”
This guide takes a focused, in-depth look at CVaR specifically — its formula, how it’s calculated using each of the standard VaR methodologies, worked examples, why it qualifies as a mathematically “coherent” risk measure in a way VaR does not, its growing role in portfolio optimization, and its adoption in banking regulation such as Basel III.
Key Takeaways
- CVaR measures the average loss within the tail of outcomes beyond the VaR threshold, directly quantifying tail severity where VaR only identifies the threshold itself.
- CVaR is always equal to or greater than VaR at the same confidence level, since it’s averaging losses that are, by definition, at least as large as the VaR threshold.
- CVaR can be calculated using historical, parametric, or Monte Carlo methods, mirroring the same approaches used for standard VaR.
- CVaR is a mathematically “coherent” risk measure, satisfying properties — including subadditivity — that standard VaR does not always satisfy.
- CVaR is more directly usable in portfolio optimization than VaR, since it can be formulated as a convex optimization problem with more favorable mathematical properties.
- Bank regulatory frameworks, including Basel III’s market risk standards, have shifted toward requiring Expected Shortfall in place of, or alongside, standard VaR.
- Like VaR, CVaR remains an estimate dependent on model assumptions and historical data, and is not a guarantee of maximum possible loss.
What Is Conditional Value at Risk?
Conditional Value at Risk is the expected (average) loss, conditional on the loss already exceeding the VaR threshold at a given confidence level. If a 95% VaR identifies the loss level that will be exceeded only 5% of the time, the corresponding 95% CVaR calculates the average loss specifically within that worst 5% of outcomes — not just where that worst-case region begins, but how severe it typically is once you’re inside it.
This distinction matters enormously in practice. Two portfolios can have an identical 95% VaR figure while having very different 95% CVaR figures — one might have a tail that, while it begins at the same threshold as the other, extends to only modestly larger losses, while the other’s tail extends to catastrophic losses far beyond that same starting threshold. VaR alone cannot distinguish between these two very different risk profiles; CVaR can.
The CVaR Formula
CVaR = E[Loss | Loss > VaR]
In words: CVaR is the expected value (average) of the loss, given that the loss exceeds the VaR threshold. This conditional expectation is what gives Conditional VaR its name — it’s the average outcome conditional on already being in the worst-case tail defined by VaR.
How CVaR Is Calculated
Historical Simulation Approach
Using the historical simulation method, CVaR is calculated by first identifying the VaR threshold (for example, the loss at the 5th percentile of historical returns for a 95% confidence level), and then simply averaging all the historical losses that were worse than that threshold — that is, averaging the actual observed losses across the entire worst 5% of historical outcomes, rather than stopping at the boundary itself.
Parametric Approach
Under the parametric method’s assumption of a normal distribution, CVaR can be calculated using a closed-form formula involving the standard normal probability density function evaluated at the relevant z-score, divided by the tail probability, then scaled by the portfolio’s standard deviation. This produces a specific multiplier, generally larger than the z-score used for standard parametric VaR at the same confidence level, reflecting the fact that CVaR captures the full tail rather than just its starting point.
Monte Carlo Approach
Using Monte Carlo simulation, CVaR is calculated by generating a large number of simulated outcomes, identifying the VaR threshold among those simulated results, and then averaging all the simulated losses that fell beyond that threshold — directly analogous to the historical simulation approach, but applied to simulated rather than actual historical data.
A Worked Example
Consider a simplified historical simulation example using 100 historical daily return observations for a $10 million portfolio, sorted from worst to best. For a 95% confidence level, the VaR threshold is identified at the 5th percentile — meaning the 5 worst observations out of 100 fall beyond this threshold.
Suppose those 5 worst historical daily losses were: $280,000, $310,000, $340,000, $410,000, and $650,000. The 95% VaR would be identified as the smallest of these, $280,000 (the threshold that’s exceeded exactly 5% of the time). The 95% CVaR, by contrast, is the average of all five of these worst-case losses:
CVaR = ($280,000 + $310,000 + $340,000 + $410,000 + $650,000) / 5 = $398,000
This example illustrates the core distinction clearly: the VaR figure of $280,000 tells you the threshold that 95% of outcomes stayed within, but says nothing about the fact that the worst-case tail actually averaged $398,000, with one outcome reaching as high as $650,000. An investor relying on the VaR figure alone would have a meaningfully understated sense of how bad the worst-case scenarios in this dataset actually were.
Why CVaR Is Always Greater Than or Equal to VaR
By definition, CVaR at a given confidence level is always mathematically greater than or equal to VaR at that same confidence level, since CVaR is the average of a set of losses that are all, individually, at least as large as the VaR threshold itself. The only scenario where they’d be exactly equal is the unusual case where every loss within the tail is identical to the threshold — in virtually all real-world cases, CVaR will be a strictly larger, more conservative figure than VaR.
This relationship is a useful sanity check: if a reported CVaR figure is smaller than the corresponding VaR figure at the same confidence level, something has gone wrong in the calculation.
CVaR as a Coherent Risk Measure
In risk management theory, a risk measure is considered coherent if it satisfies several specific mathematical properties, one of the most important being subadditivity: the risk of a combined portfolio should never be greater than the sum of the risks of its individual components measured separately. In plain terms, diversification should never make a portfolio’s measured risk worse.
Standard VaR can, in certain circumstances, violate subadditivity — producing situations where combining two portfolios into one appears, according to the VaR calculation, to create more risk than the two separate VaR figures would suggest, which contradicts the basic intuition that diversification should help, not hurt. CVaR, by contrast, satisfies subadditivity and the other properties required to be classified as a coherent risk measure under the standard mathematical framework used in risk management theory.
This theoretical advantage is a significant part of why CVaR has gained increasing favor among risk management academics and, increasingly, practitioners and regulators, despite VaR’s continued widespread use and familiarity.
CVaR in Portfolio Optimization
Beyond its role as a risk-reporting metric, CVaR has a significant practical advantage over VaR in portfolio optimization specifically: CVaR can be formulated as a convex optimization problem, which is generally well-behaved and computationally tractable to solve using standard optimization techniques. Standard VaR, by contrast, is non-convex and non-smooth in many practical formulations, making it considerably more difficult to directly incorporate into an optimization process aiming to minimize risk.
This means that when constructing a portfolio explicitly designed to minimize tail risk — rather than simply measuring the tail risk of a portfolio that’s already been constructed using other criteria — CVaR is generally the more practical and mathematically tractable choice as the risk measure to directly optimize against.
CVaR vs VaR: Side-by-Side Comparison
| Feature | Value at Risk (VaR) | Conditional VaR (CVaR) |
|---|---|---|
| What it measures | The loss threshold at a given confidence level | The average loss within the tail beyond that threshold |
| Captures tail severity? | No | Yes, directly |
| Coherent risk measure? | Not always (can violate subadditivity) | Yes, always satisfies subadditivity |
| Suitable for portfolio optimization? | Difficult; non-convex in many formulations | Well-suited; can be formulated as convex optimization |
| Relative magnitude | Smaller (or equal) at the same confidence level | Larger (or equal) at the same confidence level |
| Regulatory trend | Traditional standard, still widely used | Increasingly required or preferred (e.g., Basel III) |
CVaR and Basel III Regulatory Standards
In the years following the 2008 financial crisis, international banking regulators specifically re-examined the standard VaR-based framework that had been widely used for regulatory capital requirements. The Basel III market risk framework introduced a shift toward using Expected Shortfall (CVaR) rather than standard VaR as the primary basis for calculating regulatory capital requirements related to market risk, explicitly citing VaR’s failure to adequately capture tail risk during periods of severe market stress as a key motivation for the change.
This regulatory shift reflects the broader theoretical and practical case for CVaR discussed throughout this guide: a risk measure that captures not just where the tail begins, but how severe that tail actually is, provides more meaningful information for both individual risk management decisions and systemic financial stability considerations.
Limitations of CVaR
Still Dependent on Historical Data or Model Assumptions
Like VaR, CVaR calculated using historical simulation depends entirely on the specific historical period used, and CVaR calculated using parametric or Monte Carlo methods depends on the accuracy of the underlying distributional and correlation assumptions — none of these methods can fully anticipate a genuinely unprecedented event outside the range of the historical or modeled scenarios considered.
Requires More Data for Reliable Estimation
Because CVaR averages over the specific tail observations beyond the VaR threshold, rather than just identifying a single threshold point, it can require a larger historical sample or more simulated scenarios to produce a statistically reliable estimate, particularly at very high confidence levels where the number of qualifying tail observations becomes quite small.
Less Intuitive to Communicate Than a Single Threshold
While CVaR is more informative than VaR, a single average tail-loss figure can still obscure the actual distribution of outcomes within that tail — an extremely skewed tail, where most outcomes are moderate but a small number are catastrophic, can produce the same CVaR figure as a tail with more evenly distributed losses, even though the two scenarios carry meaningfully different practical implications.
How to Use CVaR Alongside Other Risk Metrics
- Report CVaR alongside VaR, not as a replacement in isolation, since the gap between the two figures itself communicates useful information about tail severity.
- Pair CVaR with maximum drawdown to combine a forward-looking, model-based tail-risk estimate with the actual worst historical decline a portfolio has experienced.
- Use CVaR directly in portfolio optimization when minimizing tail risk is a specific goal, given its more favorable mathematical properties compared with optimizing against VaR directly.
- Supplement CVaR with stress testing, examining specific historical or hypothetical crisis scenarios directly, rather than relying solely on statistical tail averages derived from more “normal” historical or modeled data.
- Understand the calculation method behind any reported CVaR figure — historical, parametric, or Monte Carlo — since the choice of method and underlying assumptions can meaningfully affect the resulting figure.
Frequently Asked Questions About Conditional Value at Risk
What is Conditional Value at Risk (CVaR)?
Conditional Value at Risk is the average loss a portfolio would experience specifically within its worst-case tail of outcomes, calculated as the expected loss given that the loss already exceeds the standard Value at Risk (VaR) threshold at a given confidence level.
How is CVaR different from VaR?
VaR identifies a specific loss threshold that will be exceeded only a certain percentage of the time, while CVaR calculates the average loss specifically within that worst-case tail beyond the threshold, directly capturing tail severity that VaR alone doesn’t address.
Is CVaR always larger than VaR?
Yes, at the same confidence level, CVaR is always greater than or equal to VaR, since CVaR is the average of a set of losses that are all individually at least as large as the VaR threshold itself.
What does it mean for CVaR to be a “coherent” risk measure?
A coherent risk measure satisfies specific mathematical properties, including subadditivity, which requires that combining portfolios never increases measured risk beyond the sum of their individual risks. CVaR satisfies these properties, while standard VaR can, in certain cases, violate them.
Why is CVaR used in portfolio optimization instead of VaR?
CVaR can be formulated as a convex optimization problem, making it computationally well-suited for portfolio optimization, while standard VaR is generally non-convex and non-smooth in practical formulations, making it more difficult to directly optimize against.
How does CVaR relate to Basel III banking regulations?
Basel III’s market risk framework shifted toward using Expected Shortfall (CVaR) rather than standard VaR as the primary basis for regulatory capital requirements, in direct response to VaR’s documented failure to adequately capture tail risk during the 2008 financial crisis.
Does CVaR eliminate the limitations of VaR entirely?
No. CVaR still depends on historical data or model assumptions, requires more data for reliable estimation than VaR, and while it captures tail severity better than VaR, it doesn’t reveal the full distribution of outcomes within that tail.
Final Thoughts
Conditional Value at Risk directly addresses the most consequential limitation of standard VaR: it quantifies not just where the worst-case tail begins, but how severe that tail actually is on average. Its status as a coherent risk measure, its favorable properties for portfolio optimization, and its growing role in banking regulation following the 2008 financial crisis all reflect a broader shift toward risk measures that take tail severity seriously, rather than stopping at a single threshold.
Knowing where the worst-case tail begins is useful. Knowing what that tail actually looks like, on average, is what allows a genuine risk assessment — and that’s the specific gap Conditional VaR was built to close.
