2

Suppose I have the following stationary $AR(1)$ process:

$$ y_{t}=\alpha+\rho y_{t-1} + u_{t} $$

where $u_{t} \sim \mathbb{N}(0,\sigma^{2})$, with $\sigma^{2}$ known. Suppose I have an initial condition $y_{0}$ and terminal condition $y_{T}$ and I would like to simulate my process for the periods in the interim, i.e $t = 2,\dots,T-1$. Can someone tell me what is the right distribution from which I should draw the $u_{t}$ if I want to impose both the initial and the terminal condition?

kjetil b halvorsen
  • 77,844
econ_ugrad
  • 213
  • Is that so? I am not sure about that. If I start from $y_{1}$ and simulate forward drawing $u_{t} \sim \mathcal{N}(0,\sigma^{2})$, $y_{T} \neq y_{T}^{SIM}$ which is a condition I want to impose. – econ_ugrad Jan 29 '20 at 15:55
  • I edited the question a little bit. The crucial thing is that I want to condition both on the initial and the terminal conditions. – econ_ugrad Jan 29 '20 at 15:58
  • 1
    Have you considered drawing from the (multivariate normal) conditional distribution of $Y_2,\ldots,Y_{T-1}|{Y_1,Y_T}$? – Estacionario Jan 30 '20 at 11:34

1 Answers1

1

You cannot impose both an initial and a final condition on a standard AR(1) process, as the latter is the stochastic version of a first-order difference equation whose solution is uniquely pinned down by one condition.

However, you could look for discrete versions of the Brownian Bridge process, a continuous-time process that can be parametrized so as to simultaneously satisfy an initial and a final condition. A possible result would be an auto-regressive process like the one you describe, but with time-varying coefficients: in particular both the auto-regressive coefficient $\rho$ and the variance $\sigma$ need to be zero at time $T$.

user4422
  • 1,178