## Pages

### Negative Eigenvalues in Practical Finance

Last Wednesday, I tried to focus on possible numerical difficulties when one uses a matrix that is not positive semidefinite as a correlation matrix. But is such a case of practical relevance?

Rank defect
If correlation is estimated historically and the time series of historical prices is not long enough compared to the number of assets, then the correlation matrix cannot have full rank. Assume that the time series consists of 252 observations (business days in ine year) and the universum of assets contains 300 equities. Then at least 49 (=300 - (252-1)) eigenvalues of the historic correlation matrix are zero. Numerically, these may become slightly negative.

Another source of rank defect, say, in automated trading, may occur if the same equity is traded at two different trading desks and not recognized as being identical. This case was one of my first encounters with computational finance, some twenty years ago. In a portfolio of 146 equities, 4 of them occured twice, and the customer had troubles with the Cholesky decomposition.

Wrong ansatz matrix
In chapter 22.4. of "Volatility and Correlation - The Perfect Hedger and the Fox (2nd edition)",p.691, Riccardo Rebonato writes on a two parameter ansatz matrix for correlation in a Libor market model

".... one can show that the eigenvalues of rho_{i,j}, now defined by

rho_{i,j} = exp(-beta  exp (-gamma min(T_i, T_j)) . abs(T_i - T_j))

are always all positive." Is this true?

The correlation ansatz is reasonable: Libor forward rates whose starting points are close to each other are stronger correlated, and Libor foward rates in the far future are also stronger correlated. One would expect that the Libor 12m rates starting in 19 and in 20 years are much more corelated than the ones starting in 2 and in 3 years, respectively. A typical correlation surface would then have the following shape

 Correlation surface: T_i = i, i=1,..30. beta=0.05, gamma=0.1 In this specific case, the correlation matrix is indeed postive definite.
What happens if we vary the parameters beta and gamma? Again, we use a time horizon of 30 years for annual Libor rates, and we calculate the smallest eigenvalue which should be positve.

 Smallest eigenvalue, when beta and gamma vary.
For beta = gamma = 0.5, three eigenvalues are negative, the smallest one is -0.4, the largest one 24.6. Depending on the preferred numerical method, you can run into severe troubles.

My recommendation for working with correlation matrices is to always check (e.g. by a Cholesky decomposition) if the positive definiteness is actually fulfilled. If not, use a simpler model.