VaR And Expected Shortfall
Value at Risk (VaR) and Expected Shortfall (ES) both summarize loss distributions, but they summarize different parts of the tail. VaR reports a loss threshold at a chosen confidence level, such as “the 99% VaR over one day.” ES reports the average loss given that losses exceed the VaR threshold at that same confidence level. In practice, ES responds more directly to how bad the worst outcomes are, while VaR can miss changes in tail severity beyond its cutoff.
For a concrete example, suppose a portfolio has a 99% one-day VaR of $10 million. If the loss distribution has occasional extreme crashes, ES will reflect the average of those crashes once the $10 million threshold is crossed. VaR still stays at $10 million even if the crashes become much larger, which is why ES often tracks tail risk more tightly.
Both metrics depend on the loss model, the horizon, and the confidence level. A 1-day 99% VaR from a historical simulation can differ from a 10-day 99% VaR from a parametric model, even when the underlying portfolio is unchanged. That dependency matters because tail risk is not a property of the portfolio alone; it is also a property of the modeling choices.
Common Misreads And Pain Points
People often treat VaR as a probability of ruin or a direct statement about “how likely” a catastrophic loss is. VaR is a quantile statement: at the chosen confidence level, losses are expected to be worse than the VaR threshold only with the complementary probability. That does not describe the size of losses in the tail beyond the threshold.
Another frequent mistake is mixing up confidence level with expected frequency. A 99% VaR implies about 1% of days exceed the threshold under the model, not that losses exceed the threshold once per 100 days in a guaranteed pattern. Clustering of volatility and regime shifts can create streaks, and model calibration can drift.
Tail risk measures also inherit assumptions from supporting technologies. VaR is often computed using historical simulation, parametric methods (such as assuming returns are normally distributed or using a t-distribution), or Monte Carlo simulation. Each approach encodes different beliefs about dependence between risk factors, the shape of the tail, and whether correlations stay stable under stress.
ES has its own dependency: it requires enough tail data or enough simulation depth to estimate the conditional tail mean. With limited history, ES can become noisy, and with certain parametric assumptions, ES can be overly sensitive to the chosen tail thickness. In a model review, I’ve seen teams focus on the headline number while leaving the tail estimation method under-specified—an omission that tends to surface during backtesting.
How To Choose And Use Them
Match The Metric To The Goal
Use VaR when the decision needs a single loss threshold tied to a confidence level, such as setting a preliminary risk limit or comparing portfolios on a consistent quantile basis. Use ES when the decision needs sensitivity to losses beyond the threshold, such as capital planning that reacts to tail severity. If your reporting framework already uses VaR, ES can still be added as a second lens that answers “what happens after the cutoff,” not just “where the cutoff sits.”
In risk committees, a practical pattern is to report both metrics at the same horizon and confidence level, then explain the difference in plain language. For example, “99% VaR is the threshold; 99% ES is the average loss beyond that threshold.” This reduces the chance that stakeholders interpret VaR as a tail severity measure.
Check Horizon, Confidence, And Units
Before comparing numbers, verify the horizon (one day, ten days, one month), the confidence level (95%, 97.5%, 99%), and the loss definition (gross loss, net loss, mark-to-market change). A 10-day 99% VaR is not a scaled version of a 1-day 99% VaR unless the model assumptions support that scaling. Many teams get this wrong by applying a square-root-of-time shortcut to non-linear portfolios.
As a small aside, I once reviewed a spreadsheet where the “confidence” label was correct but the horizon was silently changed from 1-day to 5-day during a copy-paste step in Excel (version 2408, if you’re curious). The resulting VaR looked “worse” and triggered a debate that ended once the horizon mismatch was found.
Validate Tail Estimation Methods
For historical simulation, confirm the lookback window and how returns are sampled. A short window can understate tail risk if it misses rare events; a long window can dilute recent regime changes. For parametric approaches, check the assumed distribution and whether the tails are heavy enough to match observed extremes. For Monte Carlo, check the number of scenarios and whether stress scenarios are included.
ES estimation is especially sensitive to tail modeling. If the model uses a distribution with finite mean in the tail, ES can be stable; if the tail is too heavy or the estimation is under-sampled, ES can swing materially. A realistic outcome target is not a single number but a stability check: rerun the calculation with slightly different lookback windows (for example, 250 vs. 500 trading days) and see whether ES changes more than VaR.
Backtest With Care, Not Just Hits
Backtesting VaR often counts exceedances (“hits”) and compares them to the expected exceedance rate. That helps detect calibration problems, but it does not measure whether the tail losses beyond VaR are getting worse. ES backtesting frameworks exist, but they require more detailed scoring rules than a simple hit count.
In practice, teams can combine exceedance counts with an additional diagnostic: compare the average loss on exceedance days to the ES implied by the model. If exceedance days are becoming more severe while hit rates remain near the expected level, VaR can look “fine” while tail risk is rising.
Educational Case Examples
Example 1: Equity portfolio with fat tails. A risk team models daily returns using historical simulation with a 250-day window. The 99% one-day VaR is $8 million, and the 99% ES is $11 million. After a market shock, the next quarter’s VaR stays near $8 million because the quantile threshold is similar, but the 99% ES rises to $15 million because the average of losses beyond the threshold increases. The committee uses ES to justify tightening risk limits even though the VaR exceedance frequency does not spike.
Example 2: Credit spread model with regime shifts. Another team uses a parametric model for changes in credit spreads and computes 97.5% VaR over ten days. VaR appears stable across months because the model’s quantile changes slowly. During a period of widening spreads, the conditional tail mean grows faster than the quantile, and ES increases more sharply. The team traces the gap to a tail-thickness parameter that was calibrated on a calmer sub-period, then updates the calibration window and re-estimates ES.
Comparison Table And Checklist
| Feature | VaR | Expected Shortfall | What To Watch |
|---|---|---|---|
| Core definition | Loss quantile at a confidence level | Average loss beyond that quantile | Whether tail severity changes after the cutoff |
| Sensitivity to extremes | Limited beyond the threshold | Directly reflects tail severity | Tail estimation noise and scenario depth |
| Backtesting focus | Exceedance frequency (“hits”) | Tail loss accuracy via scoring rules | Whether the backtest matches the metric definition |
| Common reporting risk | Misread as probability of ruin | Misread as “guaranteed average loss” | Model dependence and conditionality |
Step-by-step checklist for decision support:
- Confirm horizon and confidence level match across portfolios and time periods.
- Read the loss definition (mark-to-market change vs. realized P&L) and sign convention.
- Identify the computation method (historical, parametric, Monte Carlo) and the tail model assumptions.
- Check whether ES is computed consistently with the VaR threshold and the same confidence level.
- Run a sensitivity test: change lookback window or scenario count and observe whether ES moves more than VaR.
- Backtest exceedances for VaR and use an ES-consistent scoring approach when available.
Common Mistakes That Mislead
One recurring mistake is comparing VaR numbers computed with different confidence levels or horizons without normalization. Another is treating VaR as a measure of tail severity, then using it to justify decisions that should respond to extreme outcomes. When the tail thickens, VaR can remain stable while ES rises, which signals that the “average beyond the cutoff” has changed.
Teams also sometimes report VaR without disclosing the model method. Historical simulation and parametric VaR can produce different tail behavior even when both are labeled “99% VaR.” If the report does not state the lookback window, distributional assumptions, or scenario generation method, the number becomes hard to interpret.
ES can be misused too. Some reports present ES as if it were a guaranteed average loss, ignoring that it is conditional on exceeding the VaR threshold under the model. If the tail is poorly estimated, ES can be unstable, which shows up as large swings across recalculations. A mild frustration for analysts is that the spreadsheet may show “ES” but the underlying code version and tail settings are buried in a separate file—so the audit trail is incomplete.
FAQ
What Does A 99% VaR Mean?
A 99% VaR is the loss threshold such that, under the model, losses exceed that threshold with probability about 1% over the stated horizon and using the stated loss definition.
How Is Expected Shortfall Calculated?
Expected Shortfall at a confidence level is the conditional average loss given that losses are worse than the VaR threshold at that same confidence level, using the same horizon and loss definition.
Which Metric Better Captures Tail Risk?
Expected Shortfall captures tail severity beyond the VaR cutoff because it averages losses in the tail; VaR focuses on the cutoff quantile and does not directly reflect how large losses get after the cutoff.
Can VaR Look Stable While Tail Risk Rises?
Yes. If the quantile threshold changes slowly but the losses beyond that threshold become larger, VaR can stay similar while ES increases.
Why Do VaR And ES Depend On Model Choices?
Both metrics rely on the assumed or estimated loss distribution, including the method (historical, parametric, Monte Carlo), the lookback window or scenario design, and how dependence between risk factors is modeled.
Author's Insight
VaR and ES are both quantification tools, not direct measurements of “risk” in a universal sense. Tail risk interpretation depends on how the loss distribution is estimated and how the conditional tail mean is computed. ES tends to respond to changes in tail severity that VaR can miss, but ES can also be noisier when tail estimation is based on limited data. A careful workflow checks horizon, confidence level, loss definition, and model method before comparing results across time or portfolios.
When a report shows only one metric, the missing piece often hides in the tail. Adding ES alongside VaR, then testing sensitivity to lookback windows or scenario counts, usually reveals whether the tail is behaving differently than the quantile suggests. I’ve seen teams get more clarity by treating the two metrics as complementary rather than competing.
Key Takeaways
- VaR reports a loss threshold at a chosen confidence level; ES reports the average loss beyond that threshold.
- ES tracks tail severity more directly, while VaR can stay stable even when extreme losses worsen.
- Comparisons require matching horizon, confidence level, and loss definition, plus disclosure of the computation method.
- Backtesting should match the metric: VaR hit rates do not test tail severity, and ES needs tail-consistent evaluation.
