Cholesky Decomposition for Correlated Asset Simulation

Published: May 9, 2026

Article by Ryan O'Connell, CFA, FRM

A multi-asset Monte Carlo simulation that draws each asset’s return independently treats them as if they had zero correlation. This ignores co-movement between assets — which can materially misstate joint downside risk when assets are positively correlated, or overstate it when they move in opposite directions. Cholesky decomposition is the standard linear-algebra technique that converts a vector of independent standard normal draws into a vector of correlated draws matching a target correlation or covariance matrix.

What is Cholesky Decomposition?

Cholesky decomposition is a factorization that breaks a symmetric positive-definite matrix into the product of a lower triangular matrix and its transpose. Named after French military officer André-Louis Cholesky, it is sometimes called a “square-root” factorization because it produces a factor L such that L × L^T equals the original matrix.

The factorization requires Σ to be symmetric and positive definite — meaning all eigenvalues are strictly positive. If these conditions hold, the Cholesky factor L is unique (given the convention of positive diagonal entries).

The Cholesky Decomposition Formula

For any symmetric positive-definite matrix Σ, the Cholesky decomposition produces a unique lower triangular matrix L with positive diagonal entries:

For a 2×2 matrix, the entries of L can be computed directly:

For larger matrices, software computes L using the Cholesky-Banachiewicz or Cholesky-Crout algorithm in approximately O(n³/3) operations.

The Simulation Transform

Once you have L, generating correlated random draws is straightforward. The formula depends on whether L comes from a covariance matrix or a correlation matrix:

Where:

• z — a vector of independent standard normal draws
• μ — the vector of expected returns
• L_Σ — Cholesky factor of the covariance matrix
• L_C — Cholesky factor of the correlation matrix
• D — diagonal matrix of standard deviations, diag(σ₁, σ₂, …)

Both approaches produce the same final correlated returns. The correlation-matrix approach is often more intuitive because it separates the correlation structure from the volatility scaling. For background on estimating these matrices from historical returns, see our correlation and covariance guide.

Interpreting the Cholesky Factor

The lower triangular structure of L has a useful interpretation: each row tells you how much of each independent shock contributes to that asset’s correlated shock.

• Row 1: The first asset’s correlated shock equals 100% of the first independent draw (z₁) and nothing else. L₁₁ equals the standard deviation (for a covariance-based factor) or 1 (for a correlation-based factor).
• Row 2: The second asset’s shock is a blend of two components — a slice of the first asset’s shock (captured by L₂₁, which encodes the correlation) plus an idiosyncratic component (captured by L₂₂).
• Row k: In general, asset k’s correlated shock blends the first k independent draws. The diagonal entry L_kk represents the conditional standard deviation of asset k after accounting for the earlier factors.

Cholesky Decomposition Example

Repeating Steps 4-6 thousands of times generates a correlated joint return distribution. Equity returns of +17% tend to coincide with above-average bond returns, but with ρ = 0.30, the relationship is weak enough that plenty of paths show one asset up while the other is down — exactly the diversification dynamic you want the simulation to capture.

Covariance-matrix variant: Factoring the covariance matrix directly gives L_Σ = [[0.18, 0], [0.018, 0.0572]]. Applying μ + L_Σ × z produces the same final returns without a separate scaling step.

Cholesky Decomposition vs. Eigendecomposition

Both Cholesky and eigendecomposition can produce a “matrix square root” suitable for correlated simulation. The choice depends on whether your matrix is well-conditioned and whether you need additional capabilities like dimension reduction.

For most practitioner workflows, Cholesky is the default. Switch to eigendecomposition (or apply a “nearest-PSD” adjustment before Cholesky) when the matrix fails to factor cleanly due to estimation noise or missing data.

How to Apply Cholesky Decomposition

Here is the standard workflow for injecting correlation into a Monte Carlo simulation:

1. Estimate the correlation or covariance matrix from historical returns. Use a consistent return frequency (daily, monthly) and sufficient history to stabilize estimates. See our correlation and covariance guide for estimation methods.
2. Verify the matrix is positive definite. A quick check: attempt the Cholesky factorization — if it succeeds, the matrix is positive definite. Alternatively, confirm all eigenvalues are positive.
3. Compute the Cholesky factor L so that Σ = L × L^T. Most numerical libraries (NumPy, Excel, MATLAB) have built-in Cholesky functions.
4. Draw a vector of independent standard normals z = (z₁, z₂, …, z_n) for each simulation step.
5. Transform: X = μ + L × z (if using covariance-based L) or R = μ + D × L × z (if using correlation-based L). Feed the correlated draw into your simulation (price path, P&L, VaR percentile).

These correlated draws are the engine inside multi-asset Monte Carlo VaR and portfolio simulations. For a comparison of simulation approaches — historical, bootstrap, and Monte Carlo — see our simulation methods comparison guide.

Common Mistakes

1. Factoring a non-positive-definite matrix. Estimation noise, missing data, asynchronous returns, or pairwise correlation estimation can produce matrices that fail Cholesky. Symptoms include runtime errors or NaN values. Remedies include shrinkage estimators, nearest-PSD projection, or falling back to eigendecomposition with non-negative spectrum.

2. Mixing correlation and covariance matrix outputs. If you factor the correlation matrix, the output is standardized — you must scale by volatilities and shift by means to get returns. If you factor the covariance matrix, the output is already in return units. Confusing the two produces returns with wrong scale.

3. Regenerating IID draws per asset instead of using coherent draws. Each simulation step should use a single vector of independent draws that gets transformed by L. Generating separate random numbers for each asset destroys the correlation structure you just built.

4. Using a stale or short-window correlation matrix. Correlations estimated from a short window are noisy; correlations from a calm period may understate crisis-period co-movement. Validate your estimates against recent regimes and consider stressed correlations for tail-risk scenarios.

5. Assuming Cholesky captures tail dependence. The Cholesky transformation encodes linear correlation under multivariate normality. It does not capture asymmetric or nonlinear co-movement — assets that crash together more severely than ρ suggests. For tail-aware simulation, consider Student-t distributions or copula-based methods.

Limitations of Cholesky Decomposition

1. Requires positive definiteness. Non-positive-definite inputs (from estimation error, missing data, or asynchronous return windows) cause the factorization to fail. Common remedies include shrinkage, nearest-PSD projection, or eigendecomposition with eigenvalue clipping.

2. Captures only linear correlation under multivariate normality. Joint tail dependence — assets crashing together more than linear ρ would suggest — is not modeled. For tail-aware risk measures like Expected Shortfall, consider Student-t or copula-based simulation.

3. Sensitive to estimation error in Σ. Small errors in the input matrix propagate through the factor, especially for highly correlated assets where off-diagonal L entries dominate. Regularization and longer estimation windows help stabilize the factor.

4. Not unique under asset reordering. Different asset orderings produce different L matrices. The joint distribution is statistically identical, but specific path realizations from a fixed random seed are not directly comparable across orderings.