Comparison, in Risk & Portfolio Construction
Historical vs Parametric VaR
Both produce the same headline number, a loss threshold at a confidence level, but they get there by opposite philosophies. Historical VaR is empirical: sort the past returns and pick the quantile, letting the data speak. Parametric VaR is model-based: estimate a mean and covariance, assume a distribution, and read the quantile off the curve. The empirical route inherits whatever skew and kurtosis the sample holds; the parametric route imposes a shape and inherits that shape's blind spots. This matrix compares the tradeoffs so you can pick deliberately rather than by habit.
6 criteria. Published May 26, 2026. By AI Fin Hub Research.
Estimates VaR by taking the empirical quantile of actual historical returns over a lookback window. No distribution is assumed; the data defines the shape.
Pros
- Makes no distributional assumption, so it captures fat tails, skew, and clustering present in the sample
- Conceptually simple and easy to explain: it is literally the historical loss at that percentile
- Naturally handles nonlinear instruments if full revaluation is used, since it reprices on real scenarios
- Hard to fool yourself about the model, because there is no model beyond the chosen window
Cons
- Only as good as the window: a calm lookback yields a dangerously low VaR before a regime shift
- Needs a long, relevant history to estimate deep tail quantiles reliably
- Gives equal weight to old and recent observations unless explicitly age-weighted
- Cannot extrapolate beyond the worst loss in the sample, so it underestimates unprecedented events
Fat-tailed or nonlinear books with ample, regime-relevant history, and any case where you distrust a normal assumption
Estimates VaR from a fitted distribution, classically normal via the variance-covariance matrix, scaling the standard deviation by the confidence z-score.
Pros
- Fast and data-light: a covariance matrix scales to large portfolios cheaply
- Smooth and stable across days, since it does not jump when one extreme observation leaves the window
- Easy to stress and decompose into per-asset and per-factor risk contributions
- Can use heavier-tailed distributions, such as Student-t, to soften the normal assumption when warranted
Cons
- The normal version systematically understates tail risk because real returns are fat-tailed
- Assumes linear exposures, so it misprices options and other nonlinear payoffs unless delta-gamma adjusted
- Sensitive to the estimated covariance, which is itself noisy and time-varying
- Hides model risk: a clean number can be confidently wrong if the distribution is misspecified
Large, roughly linear portfolios needing fast daily VaR, or cases where you explicitly model a fat-tailed distribution
Decision table
See the tradeoffs side by side
| Criterion | Historical VaR | Parametric VaR |
|---|---|---|
| Distribution assumption | None, empirical quantile | Assumed, classically normal |
| Captures fat tails and skew | Yes, if in the sample | Only with a heavy-tailed distribution |
| Data requirement | Long, relevant history | Lighter, needs covariance |
| Speed at scale | Slower, especially full revaluation | Fast, matrix algebra |
| Handles nonlinear payoffs | Yes, with full revaluation | Poorly, unless delta-gamma adjusted |
| Day-to-day stability | Jumpy as extremes enter and exit window | Smooth |
Verdict
Match the method to the book and to your trust in the data. If returns are fat-tailed or the positions are nonlinear and you have a long, regime-relevant history, historical VaR is more honest because it does not impose a shape the data contradicts. If the portfolio is large and roughly linear and you need a fast, decomposable daily number, parametric VaR earns its place, but reach for a Student-t rather than a normal so the tail is not understated. Best practice is to run both and a Monte Carlo cross-check: when historical and parametric VaR disagree sharply, the disagreement is itself a warning that the distribution is misspecified or the window is unrepresentative.
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.
- 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.
- 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.
Questions people ask next
Why does normal parametric VaR understate risk?
Real asset returns have fatter tails and more frequent extreme moves than a normal distribution predicts, and they cluster in volatility. The normal model assigns these large moves a near-zero probability, so the VaR quantile sits too close to the center and the true probability of breaching it is higher than stated. Switching to a Student-t or another heavy-tailed distribution widens the tail and partly corrects this.
How much history does historical VaR need?
Enough to estimate the chosen quantile with several observations beyond it: for 99% VaR you want hundreds of days so the worst 1% is not a single noisy point. The history must also be relevant to the current regime. A two-year window of calm markets produces a low VaR that collapses the moment volatility returns, which is why many desks age-weight observations or blend in stressed periods.
Where does Monte Carlo VaR fit between these two?
Monte Carlo VaR simulates many forward scenarios from an assumed model and reads the quantile off the simulated losses. It shares the parametric method's reliance on a model but, like full-revaluation historical VaR, it can price nonlinear payoffs correctly. It is the most flexible and the most expensive of the three, and it is often used as a referee when historical and parametric estimates diverge.
Sources
- Value at Risk: The New Benchmark for Managing Financial Risk , Philippe Jorion, McGraw-Hill (2006)
- Techniques for Verifying the Accuracy of Risk Measurement Models , Paul Kupiec, Journal of Derivatives (1995)
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 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.
- 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.