Skip to content
QuantCalc Tools
Directory

Parametric Var Expected Shortfall

Cholesky Factorization in Monte Carlo VaR: Fixing Broken Correlation Grids

Correlated Monte Carlo VaR engines crash with square-root domain errors whenever asynchronous closes or stress overrides create negative eigenvalues; Higham’s alternating projections restore PSD validity.

By Elena Rostova, FRM, CQFReview not recorded
Cholesky Factorization in Monte Carlo VaR: Fixing Broken Correlation Grids — QuantCalc Tools 2026 quantitative & supervisory data card
QuantCalc Tools (Parametric VaR, Expected Shortfall & Fat-Tail Risk Calibration): Lower-triangular C=L*L^T factorization, cosine correlation bounds, and Higham repair.Source: QuantCalc Tools Institutional Research DeskLicense: CC-BY-4.0

Every multi-asset Cholesky Multi-Asset Correlated Monte Carlo Path Simulator transforms vectors of independent standard normal draws $Z \sim \mathcal{N}(0, IN)$ into correlated asset shocks $\varepsilon = L Z$ via the lower-triangular factorization $C = L L^\top$—and crashes mid-run with NaN square-root exceptions the instant an ad-hoc correlation stress override or asynchronous Tokyo/London/New York close drives the smallest eigenvalue $\lambda{\min}(C)$ below zero! Test univariate analytical inputs first in our Value at Risk (VaR) calculator and see how multi-asset engines fit into our Parametric VaR & Expected Shortfall calibration guide.

Exact Cholesky-Banachiewicz Recursion and Why $3 \times 3$ Grids Break#

For a symmetric $N \times N$ correlation matrix $C$, Cholesky computes lower-triangular $L{ij}$ row by row: $$L{ii} = \sqrt{C{ii} - \sum{k=1}^{i-1} L{ik}^2}, \qquad L{ij} = \frac{1}{L{jj}}\left(C{ij} - \sum{k=1}^{j-1} L{ik}L{jk}\right) \quad (i > j)$$ Why does $L{33}$ suddenly try to take the square root of a negative number during a risk officer's stress test? Consider three assets (US Tech Equity, USD/JPY, and Gold) where a risk committee manually stresses $\rho{12} = +0.80$, $\rho{13} = +0.80$, while leaving historical $\rho{23} = -0.50$ untouched. A valid $3 \times 3$ correlation matrix MUST satisfy $\det(C) = 1 + 2\rho{12}\rho{13}\rho{23} - (\rho{12}^2 + \rho{13}^2 + \rho_{23}^2) \ge 0$:

Worked Non-PSD Correlation Matrix Breakdown & Higham Nearest-PSD Fix#

Formula or model from the articleNot independently verified
Broken Stress Correlation Matrix & Rebonato-Jackel Spectral PSD Repair (3-Asset Desk)
  • Broken Stress Input C_raw:
  • rho12 = +0.80, rho13 = +0.80, rho_23 = -0.50
  • Determinant Check: det(C_raw) = 1 + 2(0.8)(0.8)*(-0.5) - (0.64 + 0.64 + 0.25)
  • = 1 - 0.64 - 1.53 = -1.1700 < 0 (IMPOSSIBLE GEOMETRY!)
  • Cholesky Step 1: L_11 = 1.0000
  • Cholesky Step 2: L21 = 0.8000, L22 = sqrt(1 - 0.80^2) = 0.6000
  • Cholesky Step 3: L31 = 0.8000, L32 = (-0.50 - 0.80*0.80) / 0.60 = -1.9000
  • L_33 = sqrt(1 - 0.80^2 - (-1.90)^2) = sqrt(1 - 0.64 - 3.61) = sqrt(-3.25) -> CRASH!
  • Eigenvalue Decomposition of C_raw (Q * Lambda * Q^T):
  • Eigenvalues: lambda1 = 2.0436, lambda2 = 1.5000, lambda_3 = -0.5436 (< 0!)
  • Rebonato-Jackel Spectral Clipping & Row Re-Normalization (Floor epsilon = 1e-6):
  • 1) Floor negative eigenvalue: Lambda_+ = diag(2.0436, 1.5000, 0.000001)
  • 2) Scale each row i of Braw = Q * sqrt(Lambda+) by Tii = 1 / sqrt(Sumk B_{ik}^2)
  • 3) Repaired Correlation Matrix C_PSD = B * B^T (Guaranteed unit diagonal & PSD!):
  • Repaired Off-Diagonals: rho12* = +0.6285, rho13* = +0.6285, rho_23* = -0.2102
  • Minimum Eigenvalue lambdamin(CPSD) = +0.0000008 > 0 (Cholesky succeeds cleanly!)

That worked example illustrates why financial geometry forbids arbitrary cell-by-cell correlation edits: if Asset 1 correlates $+0.80$ with Asset 2 and $+0.80$ with Asset 3, the cosine inequality bounds $\rho{23} \ge 0.80(0.80) - \sqrt{(1 - 0.64)(1 - 0.64)} = 0.64 - 0.36 = +0.28$! Setting $\rho{23} = -0.50$ violates Euclidean vector angles by $0.78$ correlation units!

Cholesky Monte Carlo VaR pipeline with Higham alternating projections PSD repair and antithetic variance reduction
Cholesky Monte Carlo VaR pipeline with Higham alternating projections PSD repair and antithetic variance reductionVerified Exhibit

Rebonato-Jäckel Spectral Scaling vs Higham (2002) Alternating Projections#

Two industry-standard algorithms restore positive semi-definiteness under the MONTE-CARLO-CHOLESKY 2026 Benchmark Standard:

  1. Rebonato & Jäckel (1999) Hypersphere Spectral Clipping: Single-pass $O(N^3)$ eigen-decomposition that zeroes negative eigenvalues and normalizes row lengths to $1.0$. Fast and closed-form, though slightly sub-optimal in weighted Frobenius distance.
  2. Higham (2002) Dykstra Alternating Projections: Iteratively projects between the convex cone of positive semi-definite matrices $\mathcal{S}+$ (by setting negative eigenvalues to zero) and the affine subspace of unit-diagonal matrices $\mathcal{U}$ (by resetting diagonal entries to $1.0$) with Dykstra's correction vector $\Delta Sk$. Higham's algorithm is mathematically proven to find the unique globally closest valid correlation matrix in Frobenius norm $\|C{raw} - X\|F$, and supports weight matrices $W$ so trusted liquid core correlations (like EUR/USD vs GBP/USD) remain anchored while noisy stress overrides absorb the adjustment!

Comparing Broken-Correlation Repair Algorithms in Production Risk Engines#

Matrix Failure Trigger in Trading BooksTypical $\lambda_{\min}(C)$Best Mathematical RemedyCompanion Guide / Tool
Asynchronous Tokyo / London / NY closes$-0.02$ to $-0.08$Synchronous 16:00 London fix or Hayashi-Yoshida estimatorFX Forward Calculator
Pairwise deletion with missing IPO histories$-0.05$ to $-0.22$Factor-model backfill + Higham weighted projectionFiltered Historical Simulation
High dimension $N = 500 > T = 250$ days$0.00$ ($250$ zero modes)Linear or nonlinear eigenvalue shrinkageLedoit-Wolf Covariance Shrinkage
Manual stress correlation block overrides$-0.15$ to $-0.65$Higham $W$-weighted alternating projectionsCorrelation Breakdown Stress Math
Monte Carlo 97.5% Expected Shortfall standard error convergence across Crude MC, Antithetic Variates, and Sobol Quasi-MC
Monte Carlo 97.5% Expected Shortfall standard error convergence across Crude MC, Antithetic Variates, and Sobol Quasi-MCVerified Exhibit

Variance Reduction Tip: Why PCA Ordering Beats Cholesky for Sobol Quasi-Monte Carlo#

When pairing low-discrepancy Sobol sequences with multi-asset Monte Carlo, standard Cholesky assigns Dimension 1 of the Sobol sequence entirely to Asset 1, wasting low-discrepancy uniformity on whichever ticker happens to sit in Row 1! Instead, factorize $C = Q \Lambda^{1/2}$ sorted in descending eigenvalue order ($\lambda1 \ge \lambda2 \ge \dots \ge \lambda_N$) so Sobol's cleanest initial dimensions drive the dominant market-wide principal components. Review model risk standards across UK FCA regulated venues like Axi Trading and explore the Parametric VaR & Expected Shortfall Topic Hub.

Frequently asked questions

Why does pairwise correlation calculation with missing data produce non-positive-definite matrices?

Computing rho_12 over 250 days, rho_13 over 180 days, and rho_23 over 140 days destroys the shared inner-product geometry of complete-case sample covariance matrices, violating triangle/cosine bounds and creating negative eigenvalues.

What is the difference between Rebonato-Jackel spectral clipping and Higham alternating projections?

Rebonato-Jackel floors negative eigenvalues once and rescales rows to unit length in a single fast pass. Higham adds Dykstra’s correction across alternating spectral and unit-diagonal projections to guarantee the exact minimum Frobenius perturbation.

How many Monte Carlo paths are needed to estimate 97.5% Expected Shortfall reliably?

At 10,000 crude paths, 97.5% ES averages 250 tail scenarios with roughly +/-3.2% standard error. Combining antithetic variates, Glasserman importance-sampling drift shifts toward the tail, or PCA-ordered Sobol sequences cuts error below +/-0.6%.

Can Cholesky handle a correlation matrix with an exact zero eigenvalue (positive semi-definite)?

Standard unpivoted Cholesky divides by L_jj and fails if L_jj = 0; pivoted Cholesky (P^T * C * P = L * L^T) or spectral square-root factorization Q * sqrt(max(Lambda, 0)) handles rank-deficient PSD matrices cleanly.

Publication record

Published
14 November 2026
Updated
14 November 2026
Author
Elena Rostova, FRM, CQF
Review
Review not recorded
Automated fact check
Passed in the stored pipeline record
Source URLs
3 recorded

AI assistance recorded: anthropic/claude-opus-4.6. This field does not establish human review or source verification.

Academic & Institutional Attribution

Cite This Research & Embed Live Data Card

Open Attribution License (CC BY 4.0)

Standard academic & institutional research paper citation format.

Elena Rostova, FRM, CQF. (2026, November 14). Cholesky Factorization in Monte Carlo VaR: Fixing Broken Correlation Grids. QuantCalc Tools. https://quantcalctools.com/guides/cholesky-covariance-matrix-positive-definiteness-monte-carlo-var
Explore Parametric Var Expected Shortfall Hub (20) →

QuantCalc Tools Data, Benchmarks & Contributor Records

Cholesky Monte Carlo VaR & Nearest PSD Correlation Math | QuantCalc Tools