VaR Methods Compared: Parametric, Historical & Monte Carlo
Choosing the right Value at Risk method can mean the difference between a risk estimate that takes milliseconds to compute and one that requires hours of simulation. It can also mean the difference between capturing your portfolio’s true tail risk and dangerously underestimating it. This guide compares the three main VaR approaches — parametric (delta-normal), historical simulation, and Monte Carlo — so you can match the right method to your portfolio’s characteristics and your organization’s computational resources.
What Are the Three VaR Methods?
VaR methods split into two fundamental approaches: local valuation uses linear (delta) approximations and assumes normal distributions for fast analytical computation. Full valuation reprices the entire portfolio under each scenario, capturing nonlinear payoffs but requiring more computation.
The parametric method uses local valuation — it approximates portfolio risk using statistical parameters and linear exposures. Historical simulation and Monte Carlo use full valuation — they reprice the portfolio under many scenarios to build an empirical distribution of potential losses.
| Method | Valuation Approach | One-Line Summary |
|---|---|---|
| Parametric (Delta-Normal) | Local | Derives VaR analytically from portfolio volatility assuming normal returns |
| Historical Simulation | Full | Reprices portfolio under actual historical risk-factor moves |
| Monte Carlo | Full | Reprices portfolio under thousands of simulated scenarios |
Each method makes different trade-offs between speed, accuracy, and flexibility. The sections below explain how each works, followed by a detailed comparison to help you choose. For step-by-step implementation details, see our dedicated guides on the parametric method, historical simulation, and Monte Carlo VaR.
Parametric (Variance-Covariance) VaR
The parametric method — also called the variance-covariance or delta-normal method — calculates VaR using a closed-form formula. It assumes portfolio returns are normally distributed and uses the portfolio’s volatility (standard deviation) to estimate potential losses.
For multi-asset portfolios, the method requires the full covariance matrix to compute portfolio volatility: σp = √(w’Σw), where w is the vector of portfolio weights and Σ is the covariance matrix of asset returns. The formula then becomes VaR = zα × σp × P.
The parametric method’s key assumptions are: (1) linear or delta-based exposures, (2) approximately joint-normal risk factor distributions, and (3) reasonably stable covariance estimates. When these hold, it delivers fast, precise VaR with easy analytical decomposition into marginal and component VaR.
The parametric method excels for diversified cash equity portfolios with primarily linear exposures. It struggles with options and other nonlinear instruments because it captures only delta exposure, missing gamma and vega risk entirely. Empirical studies by Hendricks (1996) on foreign currency portfolios found that delta-normal VaR at 99% confidence understated risk by 9–15% compared to actual tail coverage — a consequence of the normal distribution’s thin tails versus real market returns.
Historical Simulation VaR
Historical simulation is a full-valuation approach. Rather than assuming a distribution, it reprices today’s portfolio under each day’s actual historical risk-factor movements over a lookback window (typically 250 trading days).
- Collect the last T days of risk-factor returns (e.g., 250 days of stock prices, interest rates, FX rates)
- For each historical day, apply those risk-factor moves to today’s portfolio and fully reprice
- This produces T hypothetical P&L outcomes — one for each historical scenario
- Rank the outcomes from worst to best; the VaR is the loss at your chosen percentile (e.g., the 13th-worst day for 95% VaR with 250 days, since 5% of 250 = 12.5)
Because it uses actual historical returns, the method naturally captures fat tails, skewness, and nonlinear instrument behavior without requiring distributional assumptions. This makes it appealing when returns deviate significantly from normality.
However, historical simulation has limitations. It relies on a single sample path — whatever happened in your lookback window is your entire distribution. If the window omits important events (like a volatility spike), the tails will be underrepresented. Conversely, if it includes an outlier that won’t recur, that observation dominates your VaR estimate. The method is also slow to respond to structural breaks and regime changes, which analytical methods like RiskMetrics handle more easily.
At 99% confidence with a 250-day window, you have only about 2.5 observations in the tail. This leads to high sampling variation — your VaR estimate is essentially driven by just 2–3 bad days. Longer windows improve precision but may include stale data.
Monte Carlo Simulation VaR
Monte Carlo simulation is the most flexible VaR method. It generates thousands of hypothetical scenarios by simulating risk-factor paths from a specified stochastic process, then fully revalues the portfolio under each scenario.
- Specify the model: Choose a stochastic process for each risk factor (e.g., geometric Brownian motion for equities, mean-reverting process for rates) and estimate parameters (volatilities, correlations) from historical or implied data
- Simulate and reprice: Generate thousands of random price paths, fully revalue the portfolio on each path, and compile the distribution of P&L outcomes to read off the VaR percentile
Monte Carlo can incorporate nonlinear positions, non-normal distributions, time-varying volatility, implied parameters, path-dependent payoffs, and user-defined stress scenarios. As Jorion notes, it is “by far the most powerful method” for complex portfolios.
The trade-off is computational cost. A portfolio with 1,000 assets requiring 10,000 simulations means 10 million full valuations. For portfolios containing exotic options or path-dependent instruments whose valuation itself requires simulation, you face “a simulation within a simulation” — compounding the computational burden.
The Barings collapse illustrates why method choice matters. Nick Leeson held a short straddle on Nikkei futures — approximately delta-neutral because the positive delta from calls offset the negative delta from puts. A delta-normal VaR showed near-zero risk. But a full Monte Carlo simulation capturing the nonlinear payoff revealed the true 1-month 95% VaR was approximately $138 million. The parametric method’s linear approximation completely missed the risk.
Monte Carlo also carries significant model risk. Results depend entirely on the assumed stochastic process and parameters. A misspecified model produces a confident but wrong VaR estimate.
VaR Methods Compared
The table below summarizes the key trade-offs across all dimensions that matter for method selection. This comparison draws on Jorion’s Table 9-4 and empirical studies by Hendricks (1996) and Pritsker (1997).
| Dimension | Parametric (Delta-Normal) | Historical Simulation | Monte Carlo |
|---|---|---|---|
| Position Valuation | Linear (delta only) | Full revaluation | Full revaluation |
| Distribution Assumption | Normal | Empirical historical sample | User-specified (flexible) |
| Fat-Tail Handling | Poor — understates tail risk | Captures tails in recent data | Configurable via distribution choice |
| Nonlinear Instruments | Fails (delta only, misses gamma/vega) | Handles naturally via full repricing | Handles via full repricing |
| Computational Speed | Fastest (matrix multiplication) | Intermediate | Slowest (10M+ valuations typical) |
| Data Requirements | Covariance matrix only | Full historical time series | Model parameters + simulation engine |
| Calibration Burden | Low — estimate μ and Σ | Low — use raw historical data | High — specify and calibrate stochastic model |
| Regime Change Response | Fast (update covariance matrix) | Slow (old data persists in window) | Fast (re-estimate parameters) |
| Forward-Looking Scenarios | Limited | None (backward-looking only) | Flexible (user-defined scenarios) |
| Transparency | Easy to explain | Easy to explain | Difficult to explain |
| Marginal VaR Analysis | Analytical, easy | More difficult | More difficult |
| Major Pitfall | Nonlinearities, fat tails | Regime changes, unusual events | Model risk |
Speed vs. Accuracy Trade-off
Pritsker (1997) quantified this trade-off for option portfolios at 99% VaR:
- Delta method: 5.34% mean absolute error, 0.08 seconds
- Delta-gamma-MC: 3.08% error, 3.88 seconds
- Full Monte Carlo: 0% error (benchmark), 66.27 seconds
For linear portfolios, the speed advantage of parametric VaR is compelling. For options books, the accuracy gain from full valuation often justifies the computational cost.
Which VaR Method Should You Use?
Method selection depends on your portfolio’s composition, computational resources, and reporting requirements. Here’s a decision framework:
- Diversified cash equity portfolio (linear exposures): Parametric VaR — fast, accurate, and allows analytical decomposition into risk contributions
- Portfolio with moderate options overlay: Delta-gamma approximation — adds second-order accuracy at modest computational cost
- Significant options book or volatility strategies: Historical simulation or Monte Carlo — full revaluation captures gamma, vega, and nonlinear payoffs
- Path-dependent instruments (mortgages, exotics): Monte Carlo — the most flexible method for modeling path-dependence and complex payoff structures
- Need for stress scenarios or forward-looking analysis: Monte Carlo — allows user-defined extreme scenarios beyond historical experience
Industry surveys from the UK Financial Services Authority found that 42% of banks use the covariance/delta-normal approach, 31% use historical simulation, and 23% use Monte Carlo. The parametric method’s simplicity and speed make it the most common choice, though the trend has moved toward historical simulation as computing power has increased. Major institutions like JPMorgan Chase typically employ multiple methods: parametric VaR for firm-wide daily reporting and full-valuation approaches for trading desks with significant derivatives exposure.
Many institutions use multiple methods as cross-checks. Run parametric VaR for quick daily estimates, then validate weekly with historical simulation or Monte Carlo. Large discrepancies between methods often reveal model risk or unusual portfolio characteristics worth investigating.
For deeper analysis of your portfolio’s risk decomposition, see our guides on backtesting VaR and portfolio VaR risk decomposition.
Common Mistakes When Choosing a VaR Method
- Using parametric VaR for options-heavy portfolios: Delta-only approximation ignores gamma and vega risk. A delta-neutral options book can show near-zero parametric VaR while carrying substantial actual risk.
- Treating 250 days as sufficient for 99% historical VaR: With 250 days, you have only ~2.5 observations in the 99% tail. Your VaR is essentially determined by 2–3 extreme days, leading to high sampling variation. Consider longer windows or weighted schemes.
- Ignoring model risk in Monte Carlo: Results are only as good as your assumed stochastic process. A sophisticated Monte Carlo with a misspecified model can produce confident but wrong estimates.
- Not backtesting the chosen method: All methods need validation against actual P&L. If your VaR is exceeded more (or less) often than expected, the method may be miscalibrated for your portfolio.
- Using one method across all desk types: A rates desk, equity desk, and options desk have different risk profiles. Match the method to each portfolio’s characteristics rather than imposing a single firm-wide approach.
- Using too few Monte Carlo simulations: For 99% VaR, 1,000 simulations provides only ~10 tail observations. Most practitioners use 10,000+ simulations for acceptable precision.
Limitations of VaR (All Methods)
Regardless of which method you choose, VaR has fundamental limitations that apply across the board:
- VaR does not bound possible losses: Hendricks (1996) found that when VaR was exceeded in his foreign currency portfolio study, the actual loss averaged 30–40% greater than the VaR estimate. VaR tells you the threshold, not what happens beyond it.
- Fat tails remain challenging: All methods struggle with extreme events to varying degrees. Parametric VaR is most affected, but even historical simulation depends on tail events appearing in your lookback window.
- Non-subadditivity: VaR can violate the expectation that diversification reduces risk. In some cases, VaR(A+B) > VaR(A) + VaR(B). This is why regulators have moved toward Expected Shortfall (CVaR), which is subadditive.
- Model risk is universal: Parametric assumes normality, historical assumes the past represents the future, Monte Carlo assumes your stochastic model is correct. All methods embed assumptions that can fail.
VaR should be complemented by stress testing to examine specific extreme scenarios and Expected Shortfall to measure average losses beyond the VaR threshold. No single risk measure captures everything.
The “best” VaR method depends on your portfolio. Parametric VaR offers speed and analytical tractability for linear portfolios. Historical simulation provides full valuation without distributional assumptions. Monte Carlo delivers maximum flexibility for complex portfolios at the cost of computation and model risk. Most sophisticated risk management programs use multiple methods and compare results to identify blind spots.
Frequently Asked Questions
Monte Carlo is potentially the most accurate for complex portfolios — if the model is well specified. It can capture nonlinearities, fat tails, and path-dependence that other methods miss. However, accuracy depends entirely on choosing the right stochastic process and parameters. A misspecified Monte Carlo can be less accurate than a simple historical simulation.
For portfolios with primarily linear exposures (diversified equities, bonds), parametric VaR is often accurate enough and much faster. Historical simulation provides good accuracy without requiring distributional assumptions, making it a robust middle ground. The “best” method depends on your portfolio composition and your confidence in model specification.
The Basel Committee’s market risk framework has evolved significantly:
- Basel 2.5 (2009): Required VaR plus stressed VaR for regulatory capital
- FRTB (2016, revised 2019): The Fundamental Review of the Trading Book replaced VaR with Expected Shortfall (ES) at 97.5% confidence for internal models. ES better captures tail risk and is subadditive.
Under the FRTB internal-models framework, banks calculate Expected Shortfall rather than VaR for regulatory capital. However, VaR remains widely used for internal risk management, daily P&L attribution, limit monitoring, and client reporting. The three VaR methods discussed in this article are equally applicable to ES calculation — you simply read a different statistic from the same loss distribution.
The two standard confidence levels are 95% and 99%, each with different use cases:
- 95% VaR: Expected to be exceeded ~12–13 times per year with daily VaR. Provides more frequent violations for backtesting and model validation. Common for internal risk monitoring.
- 99% VaR: Expected to be exceeded ~2–3 times per year. Used historically for regulatory capital under Basel 2.5. Requires longer historical data or more simulations for statistical reliability.
Under FRTB, regulatory capital uses 97.5% Expected Shortfall rather than 99% VaR. For internal purposes, choose a confidence level that balances meaningful tail coverage with enough violations to backtest effectively. Higher confidence levels require longer lookback windows or more simulations to avoid excessive sampling variation.
Yes, and many institutions do exactly this. Common approaches include:
- Tiered approach: Use parametric VaR for quick daily estimates across the firm, then run historical simulation or Monte Carlo weekly for books with significant options exposure
- Cross-validation: Compare results across methods. Large discrepancies often reveal model risk, unusual portfolio characteristics, or data issues worth investigating
- Method by desk: Apply parametric to linear portfolios (rates, FX spot) and full valuation to nonlinear books (options, structured products)
As Jorion notes, “Perhaps the best lesson is to check VAR measures with different methodologies and then analyze the sources of differences.” Using multiple methods provides valuable cross-checks that any single method alone cannot offer.
Disclaimer
This article is for educational purposes only and does not constitute financial advice. Value at Risk is one component of a comprehensive risk management framework and should not be used in isolation. The specific figures and examples cited are illustrative and may not reflect current market conditions. Consult with qualified risk management professionals before implementing VaR methodologies for your organization.
