In Vasicek model, we have the following relation to get Discount factors given the value of short rate: $$P(t\,,T)={{e}^{A(t,T)\,-\,B(t,T){{r}_{t}}\,}}$$

So, Discount factors are known as soon as we know the short rate. But then in some references like Glasserman (pg. 115) there is a whole subsection on "Joint Simulation [of short rate] with the Discount Factor" where he talks about simulating the pair $$({r}_{t},\int_{0}^t{r(u)}du)$$.

Piterbarg's book has something similar too. So my question is - why do we need to simulate Discount factor if we have an exact analytical result.

  • 1
    $\begingroup$ Read all of page 115 in Glasserman's book. Since the process X=(r,integral of r) is jointly Gaussian, it can be simulated with precision. We simulate X instead of simulating r and then estimating the integral of r with a sum. $\endgroup$
    – AXH
    Dec 17, 2018 at 21:20
  • 1
    $\begingroup$ @AXH thats exactly my confusion - why care about integral of r when we can calculate it's exact value analytically. Also, given we can exactly simulate r too, simulation precision cannot be the reason to simulate X. But since several references do it this way, there must be a reason. $\endgroup$
    – InnocentR
    Dec 17, 2018 at 22:00

1 Answer 1


Although it's been a long time this question has been asked, I'd like to propose an answer in case someone was looking for the same thing.

First, I think there's a confusion between $P(t,T)$ and $DF(t,T)$. The former is the $t-$price of a contract paying $1$ unit of currency at date $T$ while the later is the (stochastic) discount factor at $t$ for flows occuring at $T$. The two are linked through the relationship $$ P(t,T)=\mathbb{E}^Q[DF(t,T) | \mathcal{F}_t]$$

If $r_t$ is the instantaneous short rate, then $DF(t,T)$ is given by $$ DF(t,T)=e^{-\int_t^T r_s ds}$$ and is a random variable.

Now, the argument of Glasserman is about computing $\int_t^T r_s ds$. In theory, since one has $r_t$ up to maturity on a given path, this is just a matter of doing a Riemann sum. However, this may be very "noisy" because of discretization errors. It turns ou, as AXH mentionned, that $(r_t, \int_t^T r_s ds)$ are jointly gaussian and can be simulated precisely.

  • 1
    $\begingroup$ To complete this very good answer, as usually one needs to discount quantities, it's very useful to have the $\int_0^t r_s ds$ as well as $r_t$. For example, if you are computing the CVA of an interest rates swap, you need the values of $r_t$ to value it at $t > 0$, but you also need $\int_0^t r_s ds$ to discount the exposures to today to compute your CVA value. $\endgroup$
    – byouness
    May 13, 2021 at 18:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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