VaR Model Risk Under Changing Market Conditions

Why Market Conditions Matter to VaR

A VaR estimate is conditional on a model and its inputs. Even if the calculation is correct, it can be poorly calibrated when the inputs describe an earlier market environment rather than the conditions relevant to the next risk period. A quiet estimation window, for example, may give a precise-looking volatility estimate just before volatility rises sharply. A change in the loss process can make the historical distribution itself a weak guide to current losses.

This is a form of VaR model risk: the reported quantile may not represent the loss distribution that actually applies. The issue is not limited to arithmetic. It can arise from a stale estimation window, a change in volatility, a structural break in the loss process, or a model that assumes conditions remain stable when they do not. The earlier material on backtesting provides ways to summarize forecast exceptions; here the emphasis is on diagnosing why the inputs behind a VaR forecast may have become outdated.

Definition: For a loss $X$ with cumulative distribution function $F_X$, the VaR at confidence level $\alpha$ is the $\alpha$-quantile $$ \operatorname{VaR}_{\alpha}(X)=\inf\{x:F_X(x)\geq \alpha\}. $$ When $X$ has a continuous distribution, this is a loss threshold that is exceeded with probability $1-\alpha$. A forecast VaR is calibrated only to the extent that its assumed distribution and parameters represent the loss process over the forecast horizon.

Normal VaR and Its Dependence on Inputs

Suppose the loss over the horizon of interest is modeled as $X\sim N(\mu,\sigma^2)$, where $\mu$ is the mean loss and $\sigma$ is the standard deviation of loss. Assume that the model parameters are appropriate for that horizon and that the loss distribution is normal. If $z_\alpha$ is the $\alpha$-quantile of the standard normal distribution, then

\operatorname{VaR}_{\alpha}(X)=\mu+z_\alpha\sigma.

The formula follows by standardizing $X$. Under the stated model, $Z=(X-\mu)/\sigma$ has a standard normal distribution. Thus

P(X\leq \mu+z_\alpha\sigma) =P\left(\frac{X-\mu}{\sigma}\leq z_\alpha\right) =\alpha.

Therefore, $\mu+z_\alpha\sigma$ is the $\alpha$-quantile of the modeled loss. This justification also makes the input risk clear: at a fixed confidence level, an error in the forecast mean or standard deviation directly changes the reported VaR. The formula does not apply merely because a calculation uses a normal quantile; the normal model and its parameters must be reasonable for the forecast period.

In a changing market, the standard deviation estimated from a historical window is not automatically the standard deviation that should be used for the next period. A long window may be stable but slow to react; a short window may react quickly but be more affected by sampling noise. If the loss process has shifted, neither estimate is necessarily a good forecast without further judgment or validation.

Worked Example: Comparing a Stale Window With a Volatility Scenario

A hypothetical portfolio reports one-period loss in millions of dollars. Its model assumes a zero-mean normal loss. The variance estimated from an older window is $1.8$ million dollars squared. A current volatility scenario suggests a standard deviation of $3.0$ million dollars. Compare the 99% VaR under the window estimate and the scenario estimate, and calculate the difference. Use $z_{0.99}=2.3263$.

Method. Use the normal VaR formula under each stated standard deviation. The mean is zero in both calculations, and the confidence level and horizon are held fixed so that the comparison isolates the volatility input.

Window estimate. The variance is $1.8$, so its standard deviation is

\sigma_{\text{window}}=\sqrt{1.8}=1.3416408\text{ million}.

The corresponding VaR is

\operatorname{VaR}_{0.99,\text{window}} =0+2.3263(1.3416408) =3.12105899 \approx 3.1211\text{ million}.

Scenario estimate. With standard deviation $3.0$ million dollars,

\operatorname{VaR}_{0.99,\text{scenario}} =0+2.3263(3.0) =6.9789\text{ million}.

The scenario VaR exceeds the window VaR by

6.9789-3.12105899=3.85784101 \approx 3.8578\text{ million}.

As a check, the scenario standard deviation is more than twice the window standard deviation, so the zero-mean normal VaR should also be more than twice as large. The ratio $6.9789/3.12105899$ is approximately $2.2361$, matching $3.0/1.3416408$. The common mistake is to use the variance $1.8$ as though it were a standard deviation; the correction is to take its square root before multiplying by the normal quantile.

Changing Volatility Within a Historical Window

A historical window can combine observations from different volatility regimes. A pooled estimate may describe the average variability across those regimes, but it need not describe the risk conditional on the regime currently prevailing. This distinction matters because a one-period forecast asks about the next period, not about a randomly selected day from the entire estimation window.

For example, suppose losses have mean zero in each of two regimes, with standard deviations $\sigma_1$ and $\sigma_2$, and the proportions of observations in the window are $w_1$ and $w_2$, where $w_1+w_2=1$. The variance of the pooled observations is

w_1\sigma_1^2+w_2\sigma_2^2.

This follows because the conditional means are both zero: the overall variance is the weighted average of the conditional variances. If the conditional means differ, this simple expression is incomplete; variation between the regime means also contributes to overall variance. In either case, a pooled variance is not automatically the appropriate conditional variance for a forecast made in a known current regime.

Worked Example: A Pooled Estimate During a High-Volatility Regime

A hypothetical risk team estimates daily losses in units of $100,000. In the estimation window, 80% of days are from a low-volatility regime with zero mean and standard deviation $1.0$, and 20% are from a high-volatility regime with zero mean and standard deviation $2.5$. Today’s conditions indicate that the high-volatility regime applies. Compare the 99% VaR from the pooled variance with the conditional 99% VaR for today. Use $z_{0.99}=2.3263$.

Method. First calculate the pooled variance using the regime weights and conditional variances. Then compare that estimate with a VaR using the current regime’s standard deviation. This is a variance comparison, not an average of the two regime standard deviations.

The pooled variance is

$$ 0.80(1.0^2)+0.20(2.5^2) =0.80+1.25 =2.05. $$

Therefore, the pooled standard deviation is

\sqrt{2.05}=1.4317821.

The pooled 99% VaR is

$$ 2.3263(1.4317821)=3.3308 $$

units of $100,000, or approximately $333,080. Under the current high-volatility regime, the conditional VaR is

$$ 2.3263(2.5)=5.81575 $$

units of $100,000, or $581,575. The pooled estimate is lower by $5.81575-3.3308=2.48495$ units of $100,000, approximately $248,495. As a check, because the current regime has a standard deviation greater than the pooled standard deviation, its zero-mean VaR must also be greater. The common mistake is to treat the window’s regime proportions as if they made the current day a mixture of regimes. Once the current regime is identified, the relevant forecast is conditional on that regime.

Structural Breaks and Outdated Forecasts

A structural break is a material change in the process generating losses. It may affect the mean, volatility, or distributional form. A break can result from a change in market conditions, portfolio composition, risk limits, or other factors. The important modeling question is whether the observations before the change remain informative about the loss distribution after it.

A rolling window can eventually remove pre-break observations, but it reacts only as new data enter and old data leave. A fixed window may therefore continue to include many observations from an obsolete regime. A shorter window may reduce this delay, but it can also increase estimation noise. The window length is not a purely mechanical choice: it reflects a trade-off between responsiveness and stability.

Worked Example: VaR After a Mean and Volatility Break

A hypothetical portfolio’s old one-period loss model is $X\sim N(0,2^2)$, with losses measured in millions of dollars. After a market change, a revised scenario model is $X\sim N(1.5,3^2)$. Calculate the 99% VaR under both models and estimate the probability that a loss from the revised model exceeds the old VaR. Use $z_{0.99}=2.3263$.

Method. Calculate each model’s quantile using its own mean and standard deviation. Then standardize the old VaR using the revised model to find the post-break exceedance probability.

Under the old model, the 99% VaR is

0+2.3263(2)=4.6526\text{ million}.

Under the revised model, the 99% VaR is

1.5+2.3263(3) =1.5+6.9789 =8.4789\text{ million}.

For a loss under the revised model, $X\sim N(1.5,9)$, the probability of exceeding the old VaR is

P(X>4.6526) =P\left(Z>\frac{4.6526-1.5}{3}\right) =P(Z>1.0509) \approx 0.1467.

Thus, the old threshold would be exceeded about 14.67% of the time under the revised model, rather than the 1% exceedance probability associated with a correctly specified continuous 99% VaR. The revised VaR is larger because both the mean and standard deviation increased. As a check, the old VaR lies only about $1.05$ revised-model standard deviations above the revised mean, so an exceedance probability well above 1% is plausible. The common mistake is to update volatility but retain the old mean without checking whether the location of the loss distribution also changed.

A Practical Diagnostic Process

When a VaR forecast may be relying on outdated inputs, the aim is not to assume that every recent increase in losses proves a permanent break. Instead, examine the forecast conditions systematically, compare relevant estimates, and document the assumptions behind any adjustment. A useful diagnostic sequence is:

Confirm the forecast target.
Identify the loss definition, confidence level, horizon, and portfolio being modeled. A change in any of these can make a direct comparison misleading.

Inspect the estimation window.
Check its end date, length, and whether it includes observations from materially different conditions. Note whether recent information is diluted by a long calm period.

Compare current conditions with model inputs.
Look for evidence that volatility or the loss level has changed. If the model assumes a stable distribution, assess whether that assumption remains credible.

Recalculate under explicit alternatives.
Use clearly stated scenarios, such as a higher volatility or changed mean, and quantify how much the VaR depends on those assumptions.

Monitor forecast performance and explain decisions.
Exception patterns can provide evidence that a forecast is miscalibrated, but they should be interpreted alongside the market and portfolio context. Record why inputs or windows were retained or changed.

A responsive estimate is not automatically a better estimate. If volatility is estimated from very few observations, a single unusually large loss can cause a sharp jump. Conversely, a long window can make a genuine regime change appear gradual. Scenario calculations help reveal the magnitude of this sensitivity, but they do not by themselves prove which scenario is the correct forecast.

Common Mistakes and Interpretation

Confusing variance with standard deviation. The normal VaR formula multiplies the standard deviation by $z_\alpha$, not the variance.
Assuming historical observations are equally relevant. A large sample from an obsolete regime can still be a weak guide to current risk.
Updating only one parameter. A structural break can affect the mean, volatility, or both; the model assumptions should be checked rather than changed selectively without justification.
Calling a scenario VaR a validated forecast. A scenario illustrates sensitivity to an assumption. It does not establish that the assumed conditions will occur or that the scenario model is correctly specified.
Ignoring the horizon and loss definition. Comparisons are meaningful only when the VaRs refer to compatible losses and forecast periods.

Key takeaway: VaR is a quantile of an assumed loss distribution, not a guarantee that the next period will resemble the estimation window. Diagnose whether the window and model inputs represent current conditions, quantify the sensitivity to plausible alternatives, and treat evidence of a regime change as a reason to reassess calibration.

Check Your Understanding

For each problem, keep the stated loss units and confidence level. Use the standard normal quantiles provided.

A hypothetical one-period loss is modeled as $N(0.2,0.8^2)$ million dollars. Using $z_{0.99}=2.3263$, calculate its 99% VaR and state the modeled exceedance probability. (Answer: VaR $=2.0610$ million dollars; exceedance probability $=0.0100$.)
An old zero-mean normal model uses standard deviation $1.2$ loss units, while a current volatility assessment uses $2.0$ units. At 97.5%, use $z_{0.975}=1.96$ to calculate both VaRs and their difference. (Answer: old VaR $=2.352$ units; current VaR $=3.920$ units; difference $=1.568$ units.)
A window contains zero-mean losses from two regimes: 70% have standard deviation $1.0$ and 30% have standard deviation $2.0$. Today is in the second regime. Calculate the pooled standard deviation and pooled 95% VaR, then calculate the conditional 95% VaR for today. Use $z_{0.95}=1.6449$. (Answer: pooled standard deviation $=1.3784$; pooled VaR $=2.2673$ units; current-regime VaR $=3.2898$ units.)
An old model is $N(0,1.5^2)$; a revised scenario is $N(0.5,2.5^2)$, with losses in thousands of dollars. Using $z_{0.99}=2.3263$, calculate both 99% VaRs and the standardized distance from the revised mean to the old VaR. (Answer: old VaR $=3.4895$ thousand dollars; revised VaR $=6.3158$ thousand dollars; the old VaR is $1.1958$ revised-model standard deviations above the revised mean.)
A zero-mean normal loss model uses an older estimated variance of $1.8$ million dollars squared. A current scenario assumes standard deviation $3.0$ million dollars. At 99%, use $z_{0.99}=2.3263$ to calculate both VaRs and the increase under the scenario. (Answer: window VaR $=3.1211$ million dollars; scenario VaR $=6.9789$ million dollars; increase $=3.8578$ million dollars.)

What You'll Learn