For a CIR process, which has SDE $$ dr_t = \alpha (\mu - r_t) dt + \sigma \sqrt{r_t} dW_t $$ how can I derive the increments over the discrete time-interval from $r_t$ to $r_{t+1}$?
5
-
1$\begingroup$ $dr_t$ is replaced by $\Delta r_t$, $dt$ is replaced by $\Delta t$ and $dW_t$ is replaced by $\xi_i \sqrt{\Delta t}$ where $\xi_i$ is random $N(0,1)$ $\endgroup$– nbbo2Oct 21, 2020 at 10:52
-
1$\begingroup$ An alternative, "exact" simulation would be via the non-central Chi-Squared distribution, for example explained here: deriscope.com/docs/Andersen_Jaeckel_Kahl_2010.pdf $\endgroup$– KermittfrogOct 21, 2020 at 11:27
-
1$\begingroup$ And of course quant.stackexchange.com/questions/57790/… and quant.stackexchange.com/questions/8114/… $\endgroup$– KermittfrogOct 21, 2020 at 11:28
-
$\begingroup$ @noob2, would you be able to explain why that is the case? $\endgroup$– John SmithOct 21, 2020 at 12:04
-
$\begingroup$ It is the Euler Maruyama method, the simplest (but not the only, as Kermitfrog pointed out) way to simulate an SDE in discrete time. $dW_t$ is random zero-mean and has a variance proportional to $dt$, so the standard deviation is proportional to $\sqrt{dt}$ $\endgroup$– nbbo2Oct 21, 2020 at 12:21
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
I am not totally sure I understand what you want to achieve. It seems like you are interested in discretizing CIR SDE. This can be done using the Euler-Murayama scheme for an equidistant decomposition of the time interval $[0, T]$, $\{0=t_0<\dots<t_n=T\}$.
First of all, let us write the model dynamics: $$r_t=r_0+\alpha\int_0^t(\mu-r_s)ds+\sigma\int_0^t\sqrt{r_s}dW_s$$
We need to discretize this process: $$r_{t+\Delta t}=r_t+\alpha(\mu-r_t)\Delta t+\sigma\sqrt{r_t}W_{\Delta t}$$ with $\Delta t=\frac{T}{n}$ and $W_{\Delta t}\sim\mathcal N\left(0,\frac{T}{n}\right)\Rightarrow W_{\Delta t}=\sqrt{\frac{T}{n}}\varepsilon,$ with $\varepsilon$ being a standard normal random variable.
Finally, we can use the trapezoidal rule to numerically integrate the simulated CIR rates and compute what you need (for example, the Monte Carlo zero-coupon bond prices).
