Delta-Normal VaR

The delta-normal VaR method is a parametric approach to estimating portfolio value-at-risk that assumes returns follow a normal distribution and uses linear approximation (delta) to model how portfolio value changes with market movements. It’s the simplest and fastest VaR calculation, but trades accuracy for computational speed.

Why the delta-normal method starts with normality

The delta-normal method builds on a foundational assumption: that daily (or periodic) returns across assets follow a normal (Gaussian) distribution. Under this assumption, a portfolio’s value is itself normally distributed, so you can describe its entire risk using just two parameters: the mean return and the standard deviation. This is why the method is sometimes called the variance-covariance approach — you compute the variance (or volatility) of each position and the covariances between them, then combine them into a single tail-loss estimate.

The math is elegant. If a portfolio has expected return μ and volatility σ, then a 1-day VaR at the 95% confidence level is approximately μ − 1.645 × σ (where 1.645 is the 95th percentile of the standard normal distribution). For a $10 million portfolio with 2% daily volatility, the 95% VaR is roughly $10M × 1.645 × 0.02 = $329,000. A 99% VaR uses 2.326 instead of 1.645, yielding $465,000. That one-liner captures the entire method.

Where “delta” enters: linear approximation for derivatives

For portfolios containing only stocks and bonds — assets with linear payoff functions — the normal-distribution assumption is the only approximation. But when you add options, the portfolio’s value no longer moves linearly with the underlying market. A call option has gamma (convexity), meaning its delta (price sensitivity) changes as the underlying moves. The delta-normal method ignores gamma and treats the option as if it were a linear position with a fixed delta. This is why it’s called delta-normal: it uses the option’s delta (first-order price sensitivity) and ignores the second-order (gamma) and higher-order effects.

In practice, this means delta-normal VaR works well for interest-rate swaps, currency forwards, and long stock portfolios, but systematically underestimates losses in short-option positions or during market crashes when gamma becomes extreme.

The three-step calculation

Step 1: Build the covariance matrix. Collect daily returns for all assets in the portfolio over a lookback window (usually 252 trading days for a year). Compute the correlation and volatility of each asset and pair-wise correlations. This gives you a covariance matrix Σ.

Step 2: Compute portfolio volatility. If your portfolio weights are w (a vector of position sizes), then portfolio variance is w^T × Σ × w, and volatility is the square root. This step is where the method’s speed comes from — it’s pure linear algebra, no simulation needed.

Step 3: Apply the confidence level. Multiply portfolio volatility by the appropriate quantile from the normal distribution (1.645 for 95%, 2.326 for 99%) and by the portfolio value. This gives the VaR in dollars or basis points.

Why practitioners still use it, despite its flaws

Delta-normal VaR dominated risk management from the 1990s through the 2010s. Three reasons persist:

Speed and transparency. The calculation runs in microseconds and every number is auditable. You can trace a $1 million VaR estimate back to specific asset volatilities and correlations. Historical simulation and Monte Carlo require thousands or millions of scenario recalculations. In 1995, when JPMorgan published RiskMetrics, computational power was scarce.

Regulatory acceptance. Basel III and derivative-trading regulations long treated delta-normal VaR as a baseline approach. Banks filing stress-test results often use it as a benchmark to which they compare more complex methods.

Adequate for calm markets. In quiet periods when correlations are stable and no tail events are imminent, delta-normal’s normality assumption is roughly harmless. It fails catastrophically only during crises — exactly when you most need an accurate risk estimate.

When delta-normal VaR breaks down

The method’s critical weakness emerged during the 2008 financial crisis and replayed in the 2020 COVID crash: normal distributions have no fat tails, but real market returns do. A 25-sigma (once-in-a-million-years, in normal terms) move happened on March 16, 2020, across dozens of assets simultaneously. Delta-normal VaR flagged no warning beforehand because the method is backward-looking. If the last 252 days showed calm, the covariance matrix reflects calm, and VaR estimates are complacent.

For portfolios heavy in short options (selling downside protection), delta-normal is worse than useless. A sold put position has very high gamma-at-the-money: if the market rallies 5%, you lose gamma convexity and delta-normal forecasts your loss as near zero (delta ≈ zero at the money). If the market crashes, your delta swings hard negative and losses blow through delta-normal’s estimate. The method cannot capture this gamma punch.

Delta-normal versus historical and Monte Carlo

Three major parametric and simulation VaR methods compete:

Historical simulation abandons the normal assumption entirely. It replays the last 252 days of market moves against your current portfolio. If the worst historical day caused a $500k loss, your 99% VaR is $500k. No matrix algebra, no distributional assumptions — just empirical data. But it requires unique stress scenarios to have occurred in your lookback window. A new tail risk (crypto crash, pandemics) has no historical precedent.

Monte Carlo VaR generates thousands of random paths of future returns, each obeying a specified distribution (usually multivariate normal with fitted parameters). This is computationally expensive but flexible: you can incorporate non-normal tail behavior, regime shifts, and jumps. Most models still assume multivariate normality, though, which is only marginally better than delta-normal unless you calibrate the tails explicitly.

Delta-normal wins on speed and regulatory clarity. Historical simulation wins on simplicity and tail-fidelity. Monte Carlo wins on flexibility but loses on transparency and computation.

Regulatory use today

Agencies distinguish between the standard approach (prescribed risk weights, no VaR required) and the internal models approach (banks compute their own VaR using approved methods). Under Basel III, delta-normal VaR is permitted but increasingly discouraged in favor of expected shortfall (conditional value-at-risk), which captures tail severity better. The U.S. Federal Reserve now requires certain derivatives dealers to report VaR using at least 99% confidence and 10-day holding periods. Delta-normal at 95% and 1-day horizon is treated as a bare minimum, not a standard.