I was analyzing a GARCH(1,1) process. In particular, let's say that I have a process ${y_t}$, with $t \in {1,2,...,T}$. I have created a GARCH process that can be written as:

$\sigma_t^2 = \omega + \alpha y_{t-1}^2 + \beta \sigma_{t-1}^2$,

with $t \in{1,2,...,T}$. After that, I maximize the Log Likelihood of the model, obtaining the three parameters, namely $\hat{\alpha}, \hat{\beta}, \hat{\omega}$.

Now I would like to use these estimated parameters for forecasting volatility. In particular, I can get

$\sigma_{T+1}^2 = \hat{\omega} + \hat{\alpha} y_{T}^2 + \hat{\beta} \sigma_{T}^2$

that is, I can get the estimated volatility at $(T+1)$. It is not clear in my mind how can I get, for example, $\sigma_{T+2}^2$. According to the formulation above, I shall use $y_{T+1}^2$, but my series stops at $T$. How can I get the forecasted value $y_{T+1}$?

I have found in literature that for a GARCH(1,1) and $k>2$

$\mathbb{E}_t[\sigma_{t+k}^2] = \sum_{i=0}^{k-2} (\hat{\alpha}+\hat{\beta})^i\hat{\omega} + (\hat{\alpha}+\hat{\beta})^{k-1}\sigma_{T+1}^2$

Thus I can use the forecasted value for $\sigma_{t+k}^2$ and, inverting the formulation of GARCH(1,1), get the forecasted value of $y_{T+1}$.

  • $\begingroup$ I understand that you want a point, interval or density forecast of $y_{T+1}$. You already have the forecasted variance. If you add a distributional assumption for standardized innovations and a model for the conditional mean, you will be able to obtain a density forecast for $y_{T+1}$ from which you can derive any interval or point forecast you like. $\endgroup$ Sep 6, 2019 at 11:40

1 Answer 1


Assuming you have a model of the type $y_t = \sigma_t\varepsilon_t$, with $\sigma_t^2 = \omega + \alpha y_{t-1}^2 + \beta\sigma_{t-1}^2$, that would make your $y_t$ have the martingale-difference property with respect to its history before time $T$. In other words, optimal predictions of $y_{T+h}$, for $h\geq1$, are zero. This is because $$\mathbb{E}[y_t\mid\mathcal{F}_{T-1}] = \mathbb{E}[\sigma_t\varepsilon_t\mid\mathcal{F}_{T-1}] = \sigma_t\mathbb{E}[\varepsilon_t\mid\mathcal{F}_{T-1}] = \sigma_t\mathbb{E}[\varepsilon_t] = 0$$

Now, based on your history $t\in\{1, 2,\ldots,T\}$, an approximate forecast of $y_{T+1}$ also functions as an estimate of the squared volatility at time $T+1$, given by $$ \hat{\sigma}_{T+1}^2 = \hat{\mathbb{E}}[y_{T+1}^2\mid\mathcal{F}_{T}] = \omega + \alpha y_{T}^2 + \beta\hat{\sigma}_{T}^2 $$ If you look at this as a recursive scheme for one-step ahead volatility forecasting, then you can also look at $h>1$ steps ahead with the information available at $T$. Both $y_{T+h}$ and $\sigma_{T+h}$ are random variables, and their predictions coincide: $$ \mathbb{E}[y_{T+h}^2\mid\mathcal{F}_T] = \mathbb{E}[\sigma_{T+h}^2\mid\mathcal{F}_T] = \omega + \alpha\mathbb{E}[y_{T+h-1}^2\mid\mathcal{F}_T] + \beta\mathbb{E}[\sigma_{T+h-1}^2\mid\mathcal{F}_T]$$ $$= \omega + (\alpha+\beta)\mathbb{E}[y_{T+h-1}^2\mid\mathcal{F}_T]$$ Working backwards, this leads to a general formula which only needs $y_T^2$ and $\sigma_T^2$: $$ \mathbb{E}[y_{T+h}^2\mid\mathcal{F}_T] = \omega\sum_{k=0}^{h-1}(\alpha+\beta)^k + (\alpha+\beta)^{k-1}(\alpha y_{T}^2 +\beta\sigma_T^2) $$

In my answer I have quoted heavily -- almost verbatim, in fact -- from the book Quantitative Risk Management, by McNeil, Frey, and Embrechts. Chapter 4.4.1 in particular offers a lot of good explanations that might help you better understand some of the mechanics.

  • $\begingroup$ Dear Emil, thanks for your precise answer! After my analysis, I arrive at the conclusion that I don't need the estimation of $y_{T+h}^2$, but only the ones of $\sigma$. In particular,I really appreciate your answer and help. $\endgroup$ Sep 6, 2019 at 10:58
  • $\begingroup$ Both $y_{T+h}$ and $\sigma_{T+h}$ are random variables. Hmm, in classical statistics parameters (such as $\sigma_{T+h}$) are not random variables. Moreover, predictions of $y_{T+h}$ and $\sigma_{T+h}$ do not coincide, though those of $y_{T+h}^2$ and $\sigma_{T+h}$ may. The same applies also to an approximate forecast of $y_{T+1}$ also functions as an estimate of the squared volatility at time $T+1$. $\endgroup$ Sep 6, 2019 at 11:39
  • $\begingroup$ On the other hand, I see why $\sigma_{T+h}$ can be considered a random variable given only the information set $\mathcal{F}_{T+h-2}$ or its subset. $\endgroup$ Sep 6, 2019 at 12:38
  • $\begingroup$ Please, have a look at my new post regarding the way in which I can get the Garch model volatilities assuming t-student distribution! $\endgroup$ Sep 11, 2019 at 19:31

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.