I read this definition of an ARCH(1) model:

$$r_{ŧ}=\sigma_{t|t-1}\epsilon_{t}$$ $$\sigma^{2}_{t|t-1} = \omega + \alpha r_{t-1}^{2}$$

However, when it comes to forecasting the h-step-ahead variance, I don't understand why is defined in this way. This is the h-step-ahead conditional variance:


How can I see that the equation above is correct?

Another question. I'd like to know what is going on in the first steps of the derivation of the recursive conditional variance:

$$\sigma^{2}_{t+h|t} = E(r^{2}_{t+h}|r_{t},r_{t-1},...) = E[E(\sigma^{2}_{t+h|t+h-1}\epsilon^{2}_{t+h}|r_{t+h-1},r_{t+h-2},...)|r_{t},r_{t-1},...] = \omega + \alpha \sigma^{2}_{t+h-1|t}$$

I skipped the final steps because they are easier to follow. I guess this is an instance of $E[E(X|Y)]=E(X)$ but I don't know why those conditional variables are chosen in this case.


A couple of additional questions.

  1. Based on the expression for $E(r^{2}_{t+h}|r_{t}, r_{t-1},...)$, does it mean that, for instance, $E(r^{2}_{t+h+1}|r_{t},r_{t-1}) = \sigma^{2}_{t+h+1|t}$ but since $\sigma^{2}_{t+h+1|t} = \omega + \alpha r_{t}$, then any other expectation for additional steps $t+ h + 2$, $t+h+3$, etc are going to produce the same result?

  2. When we take the expectation $E(r^{2}_{t+h}|r_{t}, r_{t-1},...)$, why don't we modify the conditional subindex from $t$ to $t+h-1$ in $\sigma^{2}_{t+h|t}$ and we do change it when we write:

$$E(r^{2}_{t+h}|r_{t},r_{t-1},...) = E[E(\sigma^{2}_{t+h|t+h-1}\epsilon^{2}_{t+h}|r_{t+h-1},r_{t+h-2},...)|r_{t},r_{t-1},...]$$


$E(r_t^2|r_{t-j},j=1,2,...)=E(\sigma^2_{t|t-1}\epsilon_t^2|r_{t-j},j=1,2,...)$. Now given $r_{t-j},j=1,2,...$, you know the values of $\sigma_{t|t-1}$ because $\sigma^2_{t|t-1}=\omega+\alpha r_{t-1}^2$. So you can take it out of expectation to have: $E(r_t^2|r_{t-j},j=1,2,...)=\sigma^2_{t|t-1}E(\epsilon_t^2|r_{t-j},j=1,2,...)$. Also $\epsilon_t$ is independent of $r_{t-j},j=1,2,...$, so (roughly speaking) any function of that is also independent of $r_{t-j},j=1,2,...$. In particular, $f(x)=x^2$. So, $\epsilon_t^2$ is independent of $r_{t-j},j=1,2,...$. Hence, $E(r_t^2|r_{t-j},j=1,2,...)=\sigma^2_{t|t-1}E(\epsilon_t^2)$. Finally note that $\epsilon_t\sim N(0,1)$. So $Var(\epsilon_t)=E(\epsilon_t^2)=1$. Therefore, $E(r_t^2|r_{t-j},j=1,2,...)=\sigma^2_{t|t-1}.$ Now we can change $t$ to $t+h$ to get $h$-step ahead forecasts, i.e. $E(r_{t+h}^2|r_{t-j},j=1,2,...)=\sigma^2_{t+h|t-1}.$ I think you can now answer your 2nd question. Hint: you condition on previous returns $r_t$ that makes your expectation easy to compute.

  • $\begingroup$ 1+. Thanks! Please, see my update to the question. I have a couple of questions I hope you can clarify. $\endgroup$ – Robert Smith Oct 22 '13 at 19:21

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.