Monte Carlo VaR For Nonlinear Exposures

Why Simulate Nonlinear Exposures?

A portfolio with options or other nonlinear positions may not gain or lose in proportion to a market-factor change. A small move can have a different effect from a large move, and the direction of the move can matter. For these portfolios, a sensitivity approximation can be useful, but a VaR estimate based only on a linear approximation may not capture the loss distribution’s shape in the tail.

Monte Carlo VaR addresses this by generating many possible market-factor outcomes and repricing the portfolio under each one. The resulting simulated losses form an empirical distribution. Its chosen quantile estimates VaR at the corresponding confidence level. The method is useful only to the extent that the simulated factors, their assumed distributions, and the repricing model are appropriate.

Definition: Let $V_0$ be the portfolio’s current value and $V_1$ its value at the risk horizon under a simulated market outcome. Define the horizon loss as $L=V_0-V_1$, so a positive value represents a loss and a negative value represents a gain. At confidence level $\alpha$, where $0<\alpha<1$, $\operatorname{VaR}_{\alpha}(L)=\inf\{l:P(L\leq l)\geq\alpha\}$. This is the $\alpha$-quantile of the loss distribution.

For $n$ equally weighted simulated losses, sort them from smallest to largest as $L_{(1)}\leq L_{(2)}\leq\cdots\leq L_{(n)}$. Using the empirical distribution’s left-continuous quantile convention, the estimated VaR is $L_{(\lceil n\alpha\rceil)}$. This convention is stated explicitly because alternative sample-quantile conventions can interpolate between order statistics and produce a different finite-sample answer.

Simulation and Repricing

A simulation begins with a model for the market factors over a specified horizon. For example, a model might generate equity returns, interest-rate changes, or exchange-rate changes. Each simulated outcome is passed through a valuation function that reflects the actual position. A call option’s value or payoff depends nonlinearly on the underlying asset price; a bond with embedded options may require repricing that reflects changes in both rates and option exercise behavior.

The essential discipline is to calculate a portfolio value for every simulated outcome, then convert that value to a loss using the same definition each time. Replacing full repricing with delta alone changes the loss model: it is not simply a faster way to calculate the same simulated VaR.

Specify the horizon and loss convention.
Define the valuation horizon, current portfolio value, and whether loss is measured as current value minus horizon value.

Specify the market-factor model.
State the factors, their simulated distributions or scenario-generation method, and any dependence assumptions used by the model.

Generate factor outcomes.
Draw $n$ outcomes from the specified model. If the model uses random draws, the simulation is subject to sampling variability.

Reprice and calculate losses.
Apply the portfolio valuation function to each outcome and calculate $L_j=V_0-V_{1,j}$ for every simulated outcome $j$.

Sort and select the quantile.
Sort the simulated losses and apply the stated empirical quantile convention to estimate VaR.

Worked Example: VaR for a Short Call

A hypothetical portfolio is short one call written on one unit of an equity index. The strike price is $100, and the premium received at the current date is $12. At the one-period horizon, assume the call’s value is its payoff $\max(S_1-100,0)$, where $S_1$ is the index level. There is no discounting over this horizon. A Monte Carlo run produces the following ten equally weighted terminal index levels: $80, $85, $90, $95, $100, $105, $110, $115, $120, and $125. Estimate the 80% VaR of the short-call loss.

Method. Reprice the call payoff at each simulated terminal index level, then calculate the short position’s loss. The initial value of the short position is $-$12$, and its horizon value is the negative of the call payoff. Thus $L=V_0-V_1=-12+\max(S_1-100,0)$.

At index levels of $80, $85, $90, $95, and $100, the payoff is $0 in each case, so the loss is $-$12$ for each scenario. At $105, $110, $115, $120, and $125, the payoffs are respectively $5, $10, $15, $20, and $25. The corresponding losses are $-$7$, $-$2$, $3, $8, and $13.

The sorted losses are $-$12, -$12, -$12, -$12, -$12, -$7, -$2, $3, $8, $13$. With $n=10$ and $\alpha=0.80$, the selected order is $\lceil 10(0.80)\rceil=8$. Therefore the empirical 80% VaR is the eighth loss, $3.

The check is that two of the ten simulated losses exceed $3, while eight are at or below $3. A common mistake is to report the call’s payoff quantile as the portfolio’s loss VaR without accounting for the premium and short position. The loss calculation above includes both.

Why Full Repricing Matters

A nonlinear position’s value is not generally a straight-line function of a market factor. For example, the payoff of a put is zero above its strike and increases as the underlying price falls below the strike. Consequently, equal-sized upward and downward factor moves need not have symmetric effects on the portfolio.

When full revaluation is feasible, the scenario valuation should reflect the specified instrument terms and valuation assumptions. A model can instead use a local approximation, such as delta and gamma, but that is an approximation to the valuation function. It may be less accurate for large changes, for positions with strong nonlinearities, or where valuation behavior changes across scenarios. Earlier material on delta and gamma and on sensitivity-based estimates provides tools for describing such approximations; Monte Carlo VaR adds a simulated loss distribution and a quantile calculation.

Worked Example: A Short Put and the Tail Losses

A hypothetical investor is short 100 put units, each on one share, with strike price $90. The investor received a premium of $4 per unit. At the one-period horizon, assume each put’s value is its payoff $\max(90-S_1,0)$, with no discounting. Eight equally weighted simulated terminal share prices are $100, $96, $92, $88, $84, $80, $76, and $72. Calculate the empirical 75% VaR of the position.

Method. Calculate the payoff per share in each scenario, multiply by 100 units to get the short position’s liability, and compare the horizon position value with its initial value. The initial value of the short position is $-100(4)=-$400$. Its horizon value is $-100\max(90-S_1,0)$. Therefore, $L=-400+100\max(90-S_1,0)$.

For terminal prices $100, $96, and $92, the put payoff is $0, so the loss is $-$400$ in each scenario. At $88, the payoff per share is $90-88=$2$, giving a total payoff of $200 and a loss of $-400+200=-$200$. At $84, the payoff is $6 per share, so the total is $600 and the loss is $200. At $80, the total payoff is $100(10)=$1,000$, so the loss is $600. At $76, the total payoff is $100(14)=$1,400$, so the loss is $1,000. At $72, the total payoff is $100(18)=$1,800$, so the loss is $1,400.

The sorted losses are $-$400, -$400, -$400, -$200, $200, $600, $1,000, $1,400$. The selected order is $\lceil 8(0.75)\rceil=6$, so the empirical 75% VaR is $600. The result is consistent with the sorted list: six of the eight losses are at or below $600.

The common mistake is to apply the premium as though it reduced each put’s payoff before multiplying by the number of units. Here the initial premium is $4 per unit, so the portfolio’s initial short-position value is $-$400$; the horizon liability is separately calculated from the full put payoff.

Monte Carlo Sampling Error

Even when the factor model and valuation function are fixed, a finite set of random draws produces a random VaR estimate. This is Monte Carlo sampling error. Increasing the number of independent draws generally reduces this error, although the rate of improvement is typically proportional to $1/\sqrt{n}$, not $1/n$.

A useful large-sample approximation is available when the loss distribution is continuous and has a positive density near its true $\alpha$-quantile. Let $q_\alpha=\operatorname{VaR}_{\alpha}(L)$, and let $f_L(q_\alpha)$ denote the loss density there. Under these conditions, the approximate standard error of the simulated quantile is

\operatorname{SE}(\widehat{\operatorname{VaR}}_{\alpha}) \approx \frac{\sqrt{\alpha(1-\alpha)/n}}{f_L(q_\alpha)}.

This approximation is not a universal guarantee: it relies on the stated regularity conditions and an appropriate loss model. In particular, a point mass or a very irregular distribution near the target quantile can make the density-based approximation unsuitable. A probability-scale interval based on the number of simulated losses below candidate values, or repeated independent simulation runs, may be more informative in such settings.

Why the approximation holds: At the true quantile $q_\alpha$, each independent simulated loss is at or below $q_\alpha$ with probability $\alpha$. The number of such losses therefore has a binomial distribution with parameters $n$ and $\alpha$, so the simulated proportion has standard deviation $\sqrt{\alpha(1-\alpha)/n}$. Near $q_\alpha$, a small change $\Delta q$ changes the cumulative probability by approximately $f_L(q_\alpha)\Delta q$. Dividing the probability-scale standard deviation by this local density gives the stated approximate standard error for the quantile.

Worked Example: Quantifying Simulation Error and Parameter Sensitivity

Suppose a hypothetical loss model is $L\sim N(0,100^2)$, with dollar amounts understood. The model parameters are treated as fixed for this calculation. A simulation uses $n=10{,}000$ independent draws to estimate the 95% VaR. Estimate the VaR and its approximate Monte Carlo standard error. Then determine the model VaR if the assumed standard deviation were $110 instead, with the same zero mean.

Step 1: Calculate the model quantile. For a normal loss with mean zero and standard deviation $100, the 95th-percentile standard normal value is approximately 1.6449. Thus the model VaR is $1.6449(100)=$164.49$, rounded to cents.

Step 2: Evaluate the density at the quantile. The standard normal density at 1.6449 is approximately 0.1031. Scaling from standard deviation 1 to $100 gives $f_L(q_{0.95})\approx0.1031/100=0.001031$ per dollar.

Step 3: Estimate the standard error. The probability-scale standard deviation is $\sqrt{0.95(0.05)/10{,}000}=\sqrt{0.00000475}\approx0.002179$. Dividing by the density gives $0.002179/0.001031\approx$2.11$. The approximate Monte Carlo standard error is therefore $2.11.

Step 4: Change the assumed parameter. With standard deviation $110, the 95% model VaR is $1.6449(110)=$180.94$, rounded to cents. This is about $16.45 higher than the VaR using standard deviation $100.

The $2.11 is an estimate of simulation error conditional on the original model and its fixed parameters. The $16.45 difference is caused by changing the assumed standard deviation, not by drawing more simulations under the original model. A common mistake is to treat a very large simulation as if it removed uncertainty about whether the distribution or its parameters were specified correctly.

Three Different Sources of Uncertainty

A reported Monte Carlo VaR should be interpreted with care. More simulated draws address only one part of the uncertainty in the result. It is useful to distinguish:

Simulation error: random variation from using a finite number of draws from a specified model. Increasing the number of independent draws can reduce this error.
Parameter uncertainty: uncertainty about estimated inputs, such as factor volatilities or correlations. Holding estimated inputs fixed during simulation does not measure this uncertainty.
Model uncertainty: uncertainty about whether the chosen distributions, dependence assumptions, horizon behavior, or valuation approach adequately represent the risks.

A simulation can therefore produce a numerically stable estimate under a model that is still a poor representation of the portfolio’s risk. Conversely, a reasonable model can produce a noisy quantile estimate if too few draws are used, especially at a high confidence level where relatively few simulated observations lie beyond the target quantile.

Common Exam Traps

Using the wrong loss sign. With $L=V_0-V_1$, a decline in portfolio value is a positive loss. Check the sign with a simple scenario before calculating the quantile.
Using an inconsistent horizon or valuation. Compare values at the same stated horizon and use the same valuation assumptions across all scenarios.
Skipping nonlinear repricing. Replacing a nonlinear valuation with delta alone can materially change tail losses. State clearly whether the calculation uses full repricing or an approximation.
Selecting the wrong order statistic. Under the convention used here, use the rank $\lceil n\alpha\rceil$ for equally weighted simulated losses. Do not select a tail observation by using $1-\alpha$ as the rank.
Confusing precision with validity. A larger $n$ reduces simulation error under the specified model; it does not validate that model or make its parameter estimates certain.

Key takeaway: Monte Carlo VaR estimates a loss quantile by simulating market factors, repricing the portfolio under each simulated outcome, and selecting an empirical loss quantile. For nonlinear exposures, the valuation step is central. Report the quantile convention and interpret simulation precision separately from parameter and model uncertainty.

Check Your Understanding

For each problem, use the stated loss convention and empirical quantile rule where applicable.

A hypothetical investor is short one call unit with strike $50 and premium received of $5. At the horizon, the call is worth $\max(S_1-50,0)$. Five equally weighted simulated prices are $45, $50, $52, $55, and $60. What is the empirical 80% VaR? (Answer: $0)
A hypothetical portfolio is short 10 put units with strike $40 and premium received of $3 per unit. At the horizon each put is worth $\max(40-S_1,0)$. Four equally weighted simulated prices are $44, $40, $38, and $34. What is the empirical 75% VaR? (Answer: -$10)
For a normal loss model with mean zero and standard deviation $80, use the 90th-percentile standard normal value 1.2816. What is the model 90% VaR, rounded to cents? (Answer: $102.53)
For $n=2{,}500$ independent simulations and $\alpha=0.99$, calculate the approximate standard deviation of the simulated proportion of losses at or below the true 99th percentile. Give the result to four decimals. (Answer: 0.0020)
A normal loss model has mean zero. At the 97.5% level use $z=1.96$. Compare its VaR for standard deviations of $90 and $105; give both VaRs and their difference to cents. (Answer: $176.40 and $205.80; difference $29.40)

What You'll Learn