Comparison, in Risk & Portfolio Construction
Monte Carlo vs Historical VaR
Both estimate a loss quantile from a set of scenarios; they differ in where the scenarios come from. Historical VaR uses the scenarios that actually happened, in the order and magnitude they happened. Monte Carlo VaR uses scenarios drawn from an assumed stochastic model, generating as many as compute allows. The first is bounded by history and free of model assumptions; the second is unbounded in coverage but only as good as the model. This matrix lays out where each is the honest choice.
6 criteria. Published May 26, 2026. By AI Fin Hub Research.
Simulates thousands of forward scenarios from an assumed return model, reprices the portfolio on each, and reads VaR off the simulated loss distribution.
Pros
- Generates scenarios the historical record never produced, covering plausible-but-unseen tail events
- Prices nonlinear and path-dependent payoffs correctly by full revaluation per path
- Lets you choose the distribution and dependence structure, including fat tails and stress correlations
- Produces a smooth, high-resolution tail because you can simulate as many paths as compute allows
Cons
- Inherits full model risk: a wrong model gives a confidently wrong VaR
- Computationally heavy, especially with full revaluation of large or exotic books
- Requires careful calibration of volatilities, correlations, and the chosen distribution
- Easy to mistake precision for accuracy, since more paths shrink sampling noise but not model error
Books with options or path dependence, stress scenarios beyond history, and any case needing tail resolution the sample cannot provide
Replays actual historical return paths over a lookback window, reprices the portfolio on each, and takes the empirical loss quantile. No model is assumed.
Pros
- No distributional or dependence model to misspecify, so it cannot be wrong about a shape it never imposes
- Preserves the real joint behavior of assets, including the way correlations spiked in past crises
- Cheap and transparent: it is the actual loss the book would have taken on real past days
- Easy to communicate and to audit, since every scenario maps to a dated historical event
Cons
- Cannot produce any loss larger than the worst event in the window, so it misses the unprecedented
- Depends entirely on the lookback: a calm window understates risk before a regime change
- Limited tail resolution because the deep tail rests on a handful of historical observations
- Gives stale or irrelevant scenarios weight unless explicitly age-weighted or filtered
Linear or simple books with long, regime-relevant history where you trust the data over any model
Decision table
See the tradeoffs side by side
| Criterion | Monte Carlo VaR | Historical VaR |
|---|---|---|
| Scenario source | Simulated from a model | Actual historical paths |
| Covers unseen events | Yes, model can generate them | No, bounded by the window |
| Model risk | High, you own the assumptions | Low, no model imposed |
| Nonlinear payoffs | Handled by per-path revaluation | Handled by per-day revaluation |
| Compute cost | High | Low to moderate |
| Tail resolution | High, arbitrarily many paths | Low, few extreme observations |
Verdict
Choose by where the risk hides. If the book is linear and the history is long and relevant, historical VaR is the more honest answer because it imposes no model that could be wrong. If the book holds options or path-dependent payoffs, or you need to stress scenarios that history never delivered, Monte Carlo is the only method that can answer, and the model risk it introduces is a price you accept knowingly. The strongest practice runs both, plus a parametric cross-check: agreement is reassuring, and a large divergence is a signal that either the historical window is unrepresentative or the simulation model is misspecified. Remember that more Monte Carlo paths reduce sampling noise, not model error.
Run the numbers next
- Playgrounds VaR Backtest — Kupiec & Christoffersen
Paste P&L + VaR series and run Kupiec POF, Christoffersen independence, and joint conditional-coverage tests. Likelihood-ratio χ² p-values.
- Generators Synthetic Market Data Generator
Generate synthetic price series — geometric Brownian motion, GARCH(1,1) with volatility clustering, regime-switching bull/bear, or copula-linked.
- Calculators Returns Distribution Analyzer
Paste a returns CSV. Histogram, normal QQ plot, skewness, excess kurtosis, Jarque-Bera test, tail-weight index. See why Sharpe alone misleads.
Questions people ask next
If I run enough Monte Carlo paths, is the VaR exact?
More paths shrink the sampling error, so the simulated quantile converges to what the model implies, but that is not the same as the true risk. The result is only as accurate as the assumed distribution, volatilities, and correlations. Running a million paths on a misspecified normal model gives a very precise estimate of the wrong number. Precision and accuracy are different, and Monte Carlo improves only the first.
Can historical VaR ever beat Monte Carlo for tail risk?
Yes, when the historical window contains genuine crises and the book is reasonably linear, because it preserves the real way correlations broke down and tails fattened, which a naive model may smooth away. Its weakness is that it cannot exceed the worst recorded loss. The complementary move, filtered historical simulation, rescales past returns by current volatility to keep historical structure while adapting to today's regime.
Which method do regulators expect?
Capital frameworks generally allow historical, parametric, or Monte Carlo approaches provided the model is backtested and validated, and many large books use Monte Carlo for portfolios rich in optionality. The supervisory emphasis is less on the method and more on backtesting the chosen model against realized exceptions and documenting the assumptions, since any of the three can pass or fail validation depending on calibration.
Sources
- Value at Risk: The New Benchmark for Managing Financial Risk , Philippe Jorion, McGraw-Hill (2006)
- Filtered Historical Simulation , Barone-Adesi, Giannopoulos, Vosper, Journal of Banking and Finance (1999)
Keep the topic connected
- Learn Value at Risk (VaR)
Value at Risk: the loss threshold you'll exceed with probability α. Why historical VaR is brittle and what it doesn't tell you about the tail.
- Learn Monte Carlo Simulation
Monte Carlo simulation in trading: when it's the right tool, when it's overkill, and the seed-discipline gotcha that ruins most published examples.
- Learn Expected Shortfall (CVaR)
Expected shortfall: the average loss given a VaR breach. Why regulators are migrating from VaR and what ES catches that VaR misses.
- Learn Volatility
Volatility as the standard deviation of returns: realized vs implied, the annualization gotcha, and why volatility-of-volatility matters.