Rebonato–Jäckel and Higham Correlation Matrix Repair
Pairwise estimates can each fall between −1 and 1 and still be incompatible as a single correlation matrix. A negative eigenvalue can imply a negative variance for some portfolios. Rebonato–Jäckel’s spectral repair clips the negative eigenvalues and rescales the matrix. Higham’s method finds the nearest valid matrix under a specified distance. The choice matters when trading desks rely on particular correlations or require a matrix that is strictly positive definite.
How it works
A correlation matrix must give a nonnegative variance for every portfolio of standardised exposures. Rebonato–Jäckel spectral repair clips negative eigenvalues and renormalises the reconstructed diagonal to one. Higham instead minimises a specified distance over valid correlation matrices. A constrained repair can preserve entries such as a required zero correlation.
Worked example
In the report's three-asset example, proposed correlations are 0.96 and 0.72 with asset A and zero between the remaining assets. The minimum eigenvalue is −0.2 and one hedge has negative implied variance. Spectral repair gives correlations about 0.813421, 0.618255 and 0.045718. For the same hedge and volatilities the resulting annual volatility is about 0.226%.
Validation and limitations
Check the observation windows and original matrix entries before fixing them. Compare how repairs change exposures of interest as well as the matrix distance. A semidefinite matrix can still be singular. If a downstream solver requires positive definiteness, document the eigenvalue margin and test the exported precision.
Further reading
- Computing the Nearest Correlation Matrix
Nicholas Higham · PDF
- Improving the Most General Methodology to Create a Valid Correlation Matrix
K. Schöttle and R. Werner · PDF