Monte Carlo VaR Estimation

From Risk-Factor Model to VaR Estimate

A portfolio’s future loss depends on the future values of the factors that drive its exposures: market prices, interest rates, claim counts, or other modeled quantities. Monte Carlo estimation turns a specified model for those factors into a collection of possible outcomes. The portfolio is revalued in each outcome, the resulting losses are collected, and the target quantile of that sample is reported as estimated VaR.

The essential discipline is to keep the stages separate. Simulating risk factors does not itself produce portfolio losses. Each simulated outcome must be passed through a valuation or loss function that reflects the exposure. Only after that step is the sample of losses ready for a quantile calculation.

Definition: For a loss random variable $X$ and confidence level $\alpha$, where $0<\alpha<1$, $\operatorname{VaR}_{\alpha}(X)$ is the lower $\alpha$-quantile: $$ \operatorname{VaR}_{\alpha}(X)=\inf\{x:F_X(x)\geq \alpha\}. $$ A Monte Carlo VaR estimate uses simulated scenarios to construct losses $X_1,\ldots,X_n$, then estimates this quantile from the empirical loss distribution.

The model must specify the distribution of each risk factor and how the factors depend on one another. For every simulated scenario, the same exposure valuation rules are applied. In a one-period portfolio setting, a common convention is $X=V_0-V_1$, where $V_0$ is current portfolio value and $V_1$ is its scenario value at the horizon. A positive $X$ is a loss; a negative $X$ is a gain. If the modeled loss is instead a future liability payment, define that loss directly and do not subtract it from an asset value unless the problem’s valuation setup calls for it.

The Monte Carlo Estimation Process

A sound estimate begins with a clearly specified horizon and valuation basis. The risk-factor model should represent outcomes over that horizon, including relevant dependence. Revaluation then translates each outcome into a loss using the same portfolio definition in every scenario. The final step applies the chosen empirical quantile convention.

Specify the model and horizon.
State the risk factors, their marginal distributions and parameters, their dependence structure, and the period over which changes are modeled.

Generate joint scenarios.
Draw a set of risk-factor outcomes from the joint model. Factors that are dependent must be simulated dependently, not as unrelated draws.

Revalue the exposures.
Apply the portfolio’s valuation or loss function to each scenario. Include nonlinear payoffs where appropriate.

Form the loss sample and extract VaR.
Order the simulated losses and select the order statistic prescribed by the empirical quantile convention.

For the convention used here, order the simulated losses as $X_{(1)}\leq X_{(2)}\leq\cdots\leq X_{(n)}$. The empirical distribution assigns mass $1/n$ to each simulated loss, counting repeated values with their multiplicities. Its lower $\alpha$-quantile is

\widehat{\operatorname{VaR}}_{\alpha}=X_{(k)},\qquad k=\lceil n\alpha\rceil.

This convention chooses the first ordered loss for which the empirical cumulative proportion is at least $\alpha$. It is important to state the convention because other software or references may use a different interpolation rule.

Why the Order-Statistic Rule Works

The empirical cumulative distribution at a value $x$ is the fraction of simulated losses no greater than $x$:

\widehat{F}_X(x)=\frac{1}{n}\sum_{j=1}^{n}\mathbf{1}(X_j\leq x).

At the $k$th ordered loss $X_{(k)}$, at least $k$ observations are no greater than that value, so $\widehat{F}_X(X_{(k)})\geq k/n$. For $k=\lceil n\alpha\rceil$, $k/n\geq\alpha$. At any value strictly below $X_{(k)}$, no more than $k-1$ observations are counted, and $(k-1)/n<\alpha$. Thus $X_{(k)}$ is the smallest value at which the empirical cumulative distribution reaches or exceeds $\alpha$, which is exactly the empirical lower-quantile definition. Ties do not invalidate the argument: they are counted as separate observations in the empirical distribution.

Worked Example: Revalue a Single Market Exposure

A hypothetical investment portfolio is worth $V_0=\$10{,}000{,}000$. Over the next year, its simple return is modeled as $R=0.03Z$, where $Z\sim N(0,1)$. There is no discounting or mortality in this one-year market-risk calculation. For eight Monte Carlo scenarios, the simulated $Z$ values are $-2,-1.5,-1,-0.5,0,0.5,1,1.5$. Estimate the 75% VaR using the stated empirical quantile convention.

Method. For each scenario, calculate the portfolio return, revalue the portfolio, and define loss as $X=V_0-V_1$. Since $V_1=V_0(1+R)$, the loss is $X=-V_0R=-\$10{,}000{,}000(0.03Z)$. Thus the loss is $-\$300{,}000Z$.

Scenario losses. Substituting the eight simulated values gives, in dollars:

$Z=-2:\ X=\$600{,}000$; $Z=-1.5:\ X=\$450{,}000$; $Z=-1:\ X=\$300{,}000$; $Z=-0.5:\ X=\$150{,}000$; $Z=0:\ X=\$0$; $Z=0.5:\ X=-\$150{,}000$; $Z=1:\ X=-\$300{,}000$; and $Z=1.5:\ X=-\$450{,}000$.

Order and select. In ascending order, the losses are $-\$450{,}000,-\$300{,}000,-\$150{,}000,\$0,\$150{,}000,\$300{,}000,\$450{,}000,\$600{,}000$. Here $n=8$ and $\alpha=0.75$, so $k=\lceil8(0.75)\rceil=\lceil6\rceil=6$. The sixth ordered loss is $\$300{,}000$.

Estimate and check. The Monte Carlo estimate is $\widehat{\operatorname{VaR}}_{0.75}=\$300{,}000$. Exactly six of the eight simulated losses are no greater than this amount, so the empirical cumulative proportion is $6/8=0.75$. The result is consistent with the selected convention. The common mistake avoided here is to report the return quantile as a loss without reversing its sign: adverse returns produce positive losses under $X=V_0-V_1$.

Revaluation Matters for Nonlinear Exposures

If a portfolio payoff is nonlinear, applying a linear approximation to risk-factor changes can omit important behavior. Monte Carlo estimation accommodates nonlinear valuation by evaluating the actual specified payoff in each scenario. The following hypothetical liability illustrates the distinction between simulating a factor and calculating the resulting loss.

Worked Example: Revalue an Indexed Benefit

A hypothetical insurer owes a one-year benefit of $L=2{,}000\max(0,100-I)$ dollars, where $I$ is a market index at the end of the year. The index model is $\ln(I/100)=0.10Z$, with $Z\sim N(0,1)$; there is no discounting or mortality in this calculation. Six simulated standard-normal values are $-1.5,-1,-0.5,0,0.5,1$. Estimate the 80% VaR of the liability payment.

Method. Transform each simulated $Z$ into an index value $I=100e^{0.10Z}$, then apply the contractual benefit formula. Because the benefit has a floor at zero, it is not valid to extend the payment linearly to index values above 100.

Transform and value. The index outcomes, rounded to two decimals for reporting, are:

$Z=-1.5:\ I=100e^{-0.15}=86.07$, so $L=2{,}000(100-86.07)=\$27{,}860$ using the displayed rounded index; $Z=-1:\ I=90.48$, so $L=2{,}000(9.52)=\$19{,}040$; $Z=-0.5:\ I=95.12$, so $L=2{,}000(4.88)=\$9{,}760$; $Z=0:\ I=100$, so $L=\$0$; $Z=0.5:\ I=105.13$, so $L=\$0$; and $Z=1:\ I=110.52$, so $L=\$0$. Values are calculated from the displayed index outcomes to the nearest dollar.

Order and select. The ordered losses are $\$0,\$0,\$0,\$9{,}760,\$19{,}040,\$27{,}860$. With $n=6$ and $\alpha=0.80$, $k=\lceil6(0.80)\rceil=\lceil4.8\rceil=5$. Therefore the estimated 80% VaR is $\$19{,}040$.

Check. Five of the six simulated payments are no greater than $\$19{,}040$, giving an empirical cumulative proportion of $5/6$, which exceeds 0.80; four of six are no greater than the preceding ordered loss, $\$9{,}760$, giving $4/6<0.80$. The selected value is therefore the empirical 80% quantile under this convention. The common mistake avoided is paying a negative benefit when the index exceeds 100; the contract’s maximum of zero must be applied in every scenario.

Joint Risk-Factor Scenarios

When more than one risk factor drives a portfolio, the joint simulation must reflect their modeled dependence. For example, if two standardized factors have correlation $\rho$, a standard construction starts with independent standard-normal variables $Z_1,Z_2$ and sets $W=\rho Z_1+\sqrt{1-\rho^2}Z_2$. Then $Z_1$ and $W$ each have standard normal distributions and correlation $\rho$. Simulating each factor independently instead would impose zero correlation and change the joint outcomes used to revalue the portfolio.

Worked Example: Combine Correlated Factor Effects

A hypothetical portfolio’s one-year change in value, in thousands of dollars, is $\Delta V=20Z_1+10W$, where $Z_1$ and $W$ are standard normal with correlation $0.6$. There is no discounting or mortality. Generate $W=0.6Z_1+0.8Z_2$ from independent standard-normal inputs. Six simulated pairs $(Z_1,Z_2)$ are $(1,0),(0,1),(-1,0),(0,-1),(-1,-1),(1,1)$. Estimate the 75% VaR, defining loss as $X=-\Delta V$.

Method. First calculate the correlated factor $W$ for each pair. Then calculate the portfolio change and reverse its sign to obtain loss. All dollar amounts below are in thousands.

Scenario calculations. For $(1,0)$, $W=0.6$, $\Delta V=20(1)+10(0.6)=26$, and $X=-26$. For $(0,1)$, $W=0.8$, $\Delta V=0+8=8$, and $X=-8$. For $(-1,0)$, $W=-0.6$, $\Delta V=-20-6=-26$, and $X=26$. For $(0,-1)$, $W=-0.8$, $\Delta V=0-8=-8$, and $X=8$. For $(-1,-1)$, $W=-0.6-0.8=-1.4$, $\Delta V=-20-14=-34$, and $X=34$. For $(1,1)$, $W=0.6+0.8=1.4$, $\Delta V=20+14=34$, and $X=-34$.

Order and select. The ordered losses are $-34,-26,-8,8,26,34$ thousand dollars. Here $n=6$, $\alpha=0.75$, and $k=\lceil6(0.75)\rceil=5$. The fifth ordered loss is $26$ thousand dollars, so $\widehat{\operatorname{VaR}}_{0.75}=\$26{,}000$.

Check. Five of the six losses are no greater than $\$26{,}000$, so the empirical proportion is $5/6\geq0.75$; only four are no greater than the preceding ordered value, $\$8{,}000$, so $4/6<0.75$. The estimate is consistent. The common mistake avoided is using independent draws for $Z_1$ and $W$, which would fail to represent the stated correlation.

Interpretation and Practical Checks

A Monte Carlo VaR estimate is conditional on the risk-factor model, the valuation rules, the horizon, and the generated scenarios. It is not a guarantee that losses cannot exceed the estimate. At confidence level $\alpha$, the quantile describes a point in the modeled loss distribution; losses beyond it remain possible.

Before reporting an estimate, check the following:

Loss sign: confirm that adverse outcomes produce positive losses under the chosen definition.
Horizon consistency: simulate factor changes over the same period used to define portfolio loss.
Dependence: preserve the assumed joint distribution when multiple factors interact.
Valuation: apply floors, caps, nonlinear payoffs, and other contractual terms in each scenario.
Quantile convention: identify the empirical rule and rank before selecting a loss.

A frequent exam trap is to sort simulated risk-factor values and take their quantile directly as portfolio VaR. That works only if the risk factor itself is the defined loss in the correct direction. In general, the portfolio’s scenario losses must first be calculated. For instance, in the first worked example, the largest return is not the largest loss: the loss is $-\$300{,}000Z$, so the worst losses arise from the smallest simulated returns.

Key takeaway: Monte Carlo VaR is obtained by simulating from a specified joint risk-factor model, revaluing the exposures scenario by scenario, forming losses with a consistent sign convention, and taking the selected empirical loss quantile. A correct quantile cannot repair a misspecified scenario model or an inconsistent valuation.

Check Your Understanding

For each problem, use the empirical convention $k=\lceil n\alpha\rceil$. All scenarios are hypothetical, and all supplied losses are already in dollars unless stated otherwise.

A one-year model produces five portfolio losses, in thousands of dollars: $4,11,7,2,9$. What is the estimated 80% VaR? (Answer: $\$9{,}000$)
A liability model produces six payments, in thousands of dollars: $0,3,3,8,12,20$. What is the estimated 75% VaR, accounting for the repeated loss? (Answer: $\$12{,}000$)
A two-factor portfolio has $\Delta V=5Z_1+4W$ in thousands of dollars, with loss $X=-\Delta V$. In one simulated scenario, $Z_1=-1$ and $W=-0.5$. Calculate the scenario loss. (Answer: $\$7{,}000$)
A hypothetical exposure has $V_0=\$2{,}000{,}000$ and scenario return $R=-0.025$. With loss defined as $V_0-V_1$, where $V_1=V_0(1+R)$, calculate the loss. (Answer: $\$50{,}000$)
Seven ordered scenario losses, in thousands of dollars, are $-5,-2,1,4,8,13,21$. Using the stated convention, identify the estimated 90% VaR and state the selected rank. (Answer: $\$21{,}000$, seventh rank)

What You'll Learn