Take the 2-minute tour ×
Quantitative Finance Stack Exchange is a question and answer site for finance professionals and academics. It's 100% free, no registration required.

I'm following loosely the exposition given in "Monte Carlo Methods in Financial Engineering by Glasserman.

For a multifactor OU process:

$dX(t)=C(b-X(t))dt+DdW(t)$

Where C and D are d*d matrices and b and X(t) are vectors on length d, and W is a d dimensional brownian motion.

He notes that this can be used to define an "exact" discretization similar to the 1D case (Shown below)

1D case:

$dr(t)=\alpha(b-r(t))dt+\sigma dW(t)$

which has solution

$r(t)=exp^{-\alpha(t-u)}r(u)+\alpha\int_u^t \exp^{- \alpha(t-s)}b(s) ds+ \sigma \int_u^t \exp^{- \alpha(t-s)}dW(s)$

and can be simulated as

$r(t+1)=exp^{-\alpha(t_{i+1}-t_i)}r(t_i)+\mu(t_i,t_{t+1})+\sigma_r(t_i,t_{i+1})Z_{i_1} $ -- EQ1

where $\mu(u,t)=\alpha \int_u^t \exp^{-\alpha(t-s)}b(s)ds$
if b is constant $\mu=b(1-\exp^{-\alpha(t_{i+1}-t_i)})$

and

$\sigma_r^2(u,t)=\sigma^2 \int_u^t \exp^{-2\alpha(t-s)}ds=\frac{\sigma^2}{2 \alpha}(1-exp^{-2 \alpha(t-u)})$

returning to multi-factor case

The general solution is

$X(t)=exp^{-C(t-u)}X(u)+\int_u^t \exp^{- C(t-s)}b ds+ \int_u^t \exp^{- C(t-s)}DdW(s)$ -- EQ2

However he also goes on to show that when C is diagonizable using $VCV^{-1}=\Delta$

$dY(t)=VX(t)$

which after simplification becomes

$dY(t)=\Delta(\tilde{b}-Y(t))dt+d \tilde W(t)$

with $\tilde{W}$ a $BM(0,\Sigma)$ where $\Sigma=VDD'T'$

This can be modeled as

$Y_j(t+1)=exp^{-\lambda_j (t_{i+1}-t_i)}Y_j(t_i)+(\exp^{\lambda_j(t_{i+1}-t_i)}-1)\tilde{b}_j +\sqrt{\frac{1}{2 \lambda_j}(1-exp^{(-2\lambda_j(t_{i+1}-t_i)})}\xi_j(i+1)$ EQ3

where $\xi(1), \xi(2)$ are independent $N(0,\Sigma)$.
Which reduces to a system of scalar simulations

You can then recover the $X(t)$ using $X(t)=V(t)^{-1}Y(t)$

My question is

1.) Could I run the multifactor OU process without doing the diagonalisation, i.e using the general solution Eq2 and the discretization scheme shown in EQ1? (if not why not?)

I realize that the second solution utilizing the diagonalisation is prolly neater (and faster) but I would like to be able to simulate the process two ways so that I can test the models against each other.

2.) If the answer to question 1 is no that how can I benchmark my model to enusre it's correct.

share|improve this question

1 Answer 1

up vote 3 down vote accepted

EQ1 is uni-variate case. EQ2 is multivariate case, in which you have to use correlated $X_t$. His way of doing is making $Y_t$ independent so that you can simulate freely. He does so by finding PC on $\Delta$. Alternatively, you could generate correlated $X_t$ in your simulation.

To benchmark your model / code, you should first test and reproduce a given example of multivariate OU and see if you match the solution. There are ways to use exact solution Multivariate exact solution , another by Gillespie, goes up to bi-variate case and here is Gillespie implemented in Matlab

share|improve this answer
    
thank you very good answer! –  Bazman May 25 at 12:12
    
Someone added then deleted the comment "Not related to the AFNSM question, hopefully." Well this is related to the NSM question I want to generate AFNSM time series to test my calibration? Would that be a problem? –  Bazman May 27 at 10:13

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.