Monte Carlo VaR Simulator

Monte Carlo VaR offers three key advantages. First, it can handle non-linear instruments (options, mortgage-backed securities, structured products) that cannot be accurately captured with delta-normal parametric methods. Second, it can model any return distribution — normal, fat-tailed (t-distribution), skewed (asymmetric), or user-specified — rather than being constrained to normality. Third, it can generate scenarios not present in historical data, covering a wider range of potential outcomes. Historical simulation is limited to what has actually happened; Monte Carlo can simulate scenarios that have not yet occurred but are possible given the modeled distribution.

The required number of simulations depends on the confidence level and desired precision. For 95% VaR: N=1,000 gives roughly 50 tail observations — barely adequate. N=10,000 gives 500 tail observations — typically sufficient for daily reporting. For 99% VaR: N=10,000 gives only 100 tail observations; N=100,000 is preferred. For 99.9% VaR (economic capital): millions of simulations may be needed. As a rule: target at least 100–200 observations in the tail for stable estimates. Variance reduction techniques can reduce computational requirements substantially.

The Cholesky decomposition factors a correlation (or covariance) matrix into a lower triangular matrix L such that LL' = Σ. It is used in Monte Carlo simulation to generate correlated random variables: if Z is a vector of independent standard normal random numbers, then R = L × Z is a vector of correlated normal random variables with covariance structure Σ. Without this step, randomly generated returns would be independent across assets, ignoring the correlations that are critical to portfolio risk. Cholesky decomposition requires the covariance matrix to be positive semi-definite — if the matrix is near-singular or has negative eigenvalues (possible with many assets), regularization techniques are needed.

As the number of risk factors (assets, yield curve points, volatility surface points) increases, the number of simulations needed to adequately cover the multivariate distribution grows rapidly. For 2 assets, a few thousand simulations cover the space reasonably well. For 100 correlated assets, even millions of simulations may leave sparse coverage of extreme joint scenarios. This computational challenge — the curse of dimensionality — is addressed through factor reduction (principal component analysis of correlation structure), stratified sampling, and quasi-random (low-discrepancy) sequences (Halton, Sobol), which cover the parameter space more uniformly than purely random sampling.

Standard Monte Carlo assumes constant volatility (σ), but actual market volatility is time-varying and mean-reverting. Stochastic volatility models (Heston model, SABR model, GARCH) model volatility as itself a random process. In Heston model Monte Carlo, both the asset price and the volatility process are simulated simultaneously using correlated Brownian motions, capturing the negative correlation between equity returns and implied volatility (the 'volatility smile'). This produces more realistic P&L distributions with negative skewness and heavier left tails, particularly important for options portfolios where vega risk (sensitivity to volatility changes) is significant.

Importance sampling is a variance reduction technique that concentrates simulations in the tail regions that matter most for VaR and CVaR estimation. Instead of sampling uniformly, the simulation probability is shifted toward more extreme scenarios, and results are reweighted to correct for this bias. The result is far more tail observations for a given number of simulations, dramatically reducing the standard error of the VaR estimate in the region of interest. Importance sampling can achieve the tail estimation accuracy of 1,000,000 standard Monte Carlo simulations with only 10,000–50,000 simulations, making it practical for institutions that need high-precision VaR with limited computational resources.

The distributional assumption is the single most important model choice in Monte Carlo VaR. Using a normal distribution underestimates tail risk because real financial returns are leptokurtic (fat-tailed) and often negatively skewed. A Student's t-distribution with 4–8 degrees of freedom typically fits equity returns better and produces 30–80% higher VaR at 99% confidence. Extreme value theory (EVT) — specifically the generalized Pareto distribution for tail modeling — is the most rigorous approach for estimating the very deep tails needed for 99.9% VaR. In practice, sophisticated risk systems often use the normal distribution for most of the return distribution but fit a separate tail model using EVT for the extreme percentiles.