Tail Risk: Portfolio Loss at 95% vs 99% VaR

Tail Risk And VaR Levels

Value at Risk (VaR) at a confidence level answers a specific question: “Over a given horizon, what loss threshold will not be exceeded with that confidence, assuming the model’s distribution holds?” A 95% VaR means losses exceed the VaR threshold about 5% of the time under the model; a 99% VaR means exceedances occur about 1% of the time. The difference between these two levels is where tail risk lives, because the 99% threshold targets rarer, more extreme outcomes.

To make this concrete, consider a one-month horizon. A 95% VaR implies roughly 1 exceedance per 20 months, while a 99% VaR implies roughly 1 exceedance per 100 months, under the model’s assumptions. Real portfolios often violate those assumptions through changing volatility, regime shifts, and liquidity stress, so the “about” matters. I often see people treat VaR as a guarantee; it is not, and the confidence level does not mean “safe.”

Tail risk is not just “bigger numbers.” It is the behavior of losses in the far end of the distribution, where correlations can spike, hedges can fail, and market depth can vanish. VaR is one lens on that tail, but it depends heavily on how the distribution is modeled and how the portfolio is revalued under shocks.

What People Get Wrong

Many readers interpret VaR as a probability of ruin or a direct measure of expected worst-case loss. VaR is a quantile of the loss distribution, not an average of losses beyond the threshold. Two portfolios can share the same 99% VaR while having very different losses when the threshold is breached.

Another common mistake is assuming that moving from 95% to 99% increases VaR by a fixed percentage. In many models, the jump depends on the tail shape. If the loss distribution has heavier tails than a normal distribution, the 99% quantile can rise disproportionately. Even with a normal model, the mapping from confidence to quantile is nonlinear, so the gap between 95% and 99% already grows with horizon and volatility.

Dependencies drive a lot of the “tail.” Correlations between assets often change during stress, and the direction of change can matter. A portfolio that looks diversified under calm conditions can become concentrated in a single risk factor during a drawdown. If your VaR engine uses static correlations, it can understate tail losses when correlations rise in downturns.

Supporting technologies also shape the result. Historical simulation uses past returns; if the sample lacks relevant stress periods, the 99% tail can be thin. Parametric VaR assumes a distribution (often normal or t-distributed) and estimates parameters; small sample errors can distort extreme quantiles. Monte Carlo VaR can model nonlinear payoffs, but it still needs assumptions about volatility dynamics and market behavior during shocks.

Liquidity and rebalancing are another dependency that gets ignored. VaR often assumes you can mark positions to mid prices and trade without market impact. In real stress, bid-ask spreads widen and market depth drops, which can worsen realized losses beyond what a price-only VaR model predicts. I once reviewed a risk report where the VaR horizon was one day, but the portfolio’s actual liquidation time was closer to a week; the mismatch made the tail look artificially contained.

How To Interpret 95% Vs 99%

Start by aligning the VaR horizon with your decision horizon. A one-day VaR answers a different question than a one-month VaR, because volatility clustering and compounding effects change the distribution. If you rebalance weekly, a daily VaR can still help, but you should translate it into a consistent horizon using the model’s scaling assumptions or by running the engine at the same horizon.

Next, check whether the VaR is computed on returns or on revaluations. Returns-based VaR can miss nonlinearities from options, credit spreads, or funding costs. Revaluation-based VaR typically uses a pricing model to map risk factor shocks into portfolio value changes, which can better capture convexity and spread sensitivity.

Then look for the model family. Parametric VaR with normality often understates tail risk for assets with fat tails. A t-distribution can improve tail fit, but it still depends on degrees of freedom estimates. Historical simulation can be sensitive to the chosen window length; a 250-day window and a 1000-day window can produce different 99% quantiles, especially if the earlier period includes a crisis regime.

Finally, interpret the exceedance frequency as model-implied, not empirical. If you backtest VaR and see more exceedances than expected, the tail is not behaving like the model assumes. Backtesting is not a one-time check; it is a diagnostic loop that tells you whether the 99% tail is credible.

Step 1: Match Horizon And Liquidity

Use a VaR horizon that matches how long you cannot exit positions. If your positions require several trading days to unwind, a one-day VaR can understate tail losses because it ignores price impact and spread widening. A practical approach is to run VaR at multiple horizons (for example, 1 day, 5 days, and 20 days) and compare how the 99% number scales.

For tools, many risk workflows use risk engines that support horizon scaling and stress scenarios. If you are using a spreadsheet-based model, document the scaling rule you apply; square-root-of-time scaling works only under restrictive assumptions. I have seen analysts apply square-root scaling to portfolios with options and credit spread exposure, which can distort the 99% tail.

Outcome expectation: if your liquidity horizon is longer than the VaR horizon, the 99% VaR you rely on for decisions should be treated as a lower bound. The gap can be modest for liquid equities and large for less liquid credit or structured products.

Step 2: Compare VaR With CVaR

VaR tells you the loss threshold; CVaR (also called Expected Shortfall) summarizes the average loss beyond that threshold. For tail risk, CVaR often matters more because it responds to how bad losses get after the breach. If 95% VaR and 99% VaR differ, CVaR can reveal whether the tail is just “wider” or “much worse” once you cross the threshold.

In practice, compute both metrics using the same model assumptions. If your 99% VaR rises sharply but CVaR rises even more, the tail is heavy and the worst outcomes dominate. If VaR rises but CVaR stays relatively stable, the tail beyond the threshold may not be as extreme as the quantile suggests, which can happen with certain distribution shapes.

Outcome expectation: CVaR gives a more decision-relevant picture for risk limits tied to capital buffers, because it reflects losses in the tail rather than only the cutoff point.

Step 3: Stress Test The Tail Drivers

Run scenario analysis that targets the risk factors most likely to move together in stress. For an equity-heavy portfolio, that can include simultaneous equity selloffs and volatility spikes. For a multi-asset portfolio, include funding stress, credit spread widening, and FX moves that historically co-occurred.

Use scenarios that reflect plausible mechanisms, not only historical dates. Historical replay can be useful, but it inherits the same limitation as historical simulation: the future may not resemble the past. A practical method is to build a “shock ladder” where you scale factor moves and observe how portfolio losses evolve at the 95% and 99% levels.

Outcome expectation: stress tests often show that the 99% VaR underestimates losses when correlations jump or when volatility dynamics are mis-modeled. When stress results exceed VaR, treat VaR as a baseline model output, not a ceiling.

Step 4: Backtest And Audit Assumptions

Backtesting compares realized losses to the VaR threshold. For 95% VaR, you expect exceedances around 5% of the time; for 99% VaR, around 1%. If you observe materially more exceedances, the tail is not captured well, and the 99% number becomes less trustworthy.

Audit the inputs: return data quality, corporate actions handling, pricing model calibration, and the treatment of dividends and carry. A small data issue can distort the tail quantile more than the median. I once saw a risk team exclude a small set of days with missing prices; the resulting gap reduced the frequency of extreme moves in the historical window, which made the 99% VaR look calmer than it should have been.

Outcome expectation: a credible 99% VaR typically comes with documented backtesting results and a clear explanation of model limitations.

Case Examples With Realistic Constraints

Example 1: Equity And Bonds With Static Correlations

An investor holds a diversified mix of large-cap equities and intermediate-term government bonds. The risk model uses parametric VaR with a normal distribution and static correlations estimated from the last 12 months. During a stress period, equity volatility rises and correlations between equities and rates shift toward the same direction, reducing the diversification benefit.

The 95% VaR looks reasonable compared with realized losses, but the 99% VaR underestimates the drawdown. Backtesting shows more exceedances than the 1% target for the 99% level. A stress test that forces a joint equity selloff and volatility spike produces losses closer to the realized tail, suggesting the correlation and volatility dynamics were the weak link.

Example 2: Option-Heavy Portfolio With Liquidity Friction

A portfolio includes short-dated options and a small allocation to credit-sensitive instruments. The VaR engine revalues positions using a pricing model, but the model assumes you can transact at mid prices and that implied volatility changes follow the calibration used in the engine. In a market dislocation, bid-ask spreads widen and implied volatility moves faster than the model’s parameter update.

The 95% VaR does not breach often, but the 99% tail shows larger realized losses than the VaR threshold. CVaR computed from the model also underestimates the average loss beyond the 99% cutoff. A scenario analysis that adds spread widening and a volatility jump aligned with observed market behavior produces a tail loss distribution that better matches the realized outcomes.

VaR Level Comparison Checklist

The table below compares how 95% and 99% VaR typically behave in practice. It is decision support, not a guarantee of outcomes.

Step-by-step checklist for comparing 95% vs 99% VaR in your own reporting:

1. Confirm the horizon and the revaluation method used for VaR.
2. Record the model family (historical, parametric, or Monte Carlo) and the distribution assumptions.
3. Compute or request CVaR at the same confidence level and compare tail behavior beyond the cutoff.
4. Run stress scenarios that target joint factor moves and liquidity friction, then compare scenario losses to both VaR levels.
5. Backtest both VaR levels over a window long enough to observe tail exceedances, then review exceedance clustering.

Common Mistakes In Risk Reporting

One mistake is presenting 99% VaR without context on the model and horizon. A 99% number computed on a one-day horizon can look “small” compared with a one-month horizon, even when the underlying tail risk is similar. Readers need the horizon and the revaluation method to interpret the magnitude.

Another mistake is mixing metrics computed under different assumptions. Comparing a historical-simulation 99% VaR from one system to a parametric 99% VaR from another system can produce differences driven by methodology rather than portfolio risk. If you must compare, normalize by using consistent horizons, pricing models, and risk factor definitions.

Some reports hide the tail by reporting only VaR and not CVaR or stress results. VaR can look stable while losses beyond the cutoff worsen, especially for portfolios with nonlinear payoffs. A mild frustration: many dashboards show a single line item, and the tail story stays trapped in footnotes.

Finally, avoid treating VaR as a guarantee for regulatory or internal limits. Even when VaR is computed carefully, it remains a model-based quantile. If the model assumptions fail during stress, the realized losses can exceed the VaR threshold.

FAQ

What Does 95% VaR Mean?

95% VaR is the loss level that the model expects to not be exceeded about 95% of the time over the chosen horizon. Exceedances occur about 5% of the time under the model, not “never.”

Why Is 99% VaR Often Much Larger?

The 99% level targets rarer outcomes in the far tail, where distribution shape and parameter estimation errors matter more. Correlation and volatility shifts during stress can also widen the tail beyond what a simple model assumes.

Is VaR A Worst-Case Loss?

No. VaR is a quantile threshold, not the maximum loss. Losses beyond the VaR cutoff can be much larger, which is why CVaR or stress tests often matter for tail risk decisions.

How Should I Compare VaR Across Portfolios?

Compare VaR only when horizons, revaluation methods, and model assumptions match. If one portfolio uses historical simulation and another uses parametric VaR, the 95% and 99% numbers can differ for methodological reasons.

What Backtesting Tells You About Tail Risk?

Backtesting counts exceedances versus the expected frequency for each confidence level. If 99% VaR shows more exceedances than expected or if exceedances cluster in certain regimes, the tail model fit is weak.

Author's Insight

VaR at 95% versus 99% is a practical way to probe tail behavior, but it depends on horizon, model family, and how the portfolio is revalued under shocks. In many real portfolios, the biggest driver of the 99% tail gap is not the confidence level alone; it is changes in correlations, volatility dynamics, and liquidity assumptions. A careful workflow pairs VaR with CVaR and scenario analysis, then checks backtesting results for both confidence levels. If you see a large 99% VaR jump, treat it as a signal to audit tail assumptions rather than as a reason to ignore the model.

Key Takeaways

• 95% VaR and 99% VaR are quantiles; the 99% level targets rarer tail outcomes and can rise disproportionately.
• VaR is model-based and horizon-specific; it does not guarantee worst-case losses or trading outcomes under liquidity stress.
• Tail risk interpretation improves when you compare VaR with CVaR and run stress scenarios that reflect joint factor moves.
• Backtesting exceedances and auditing model assumptions determine whether the 99% tail number is credible for decisions.