Suppose a monthly, stationary time series. The series seems to have some ARCH effects and I model its variance as a GARCH process. I obtain the following output of a GARCH(1,1) model:

  • alpha ($\alpha$): 0,07
  • beta ($\beta$): 0,7
  • intercept ($c$): 0,00008

I obtain the unconditional variance of the GARCH(1,1) model by taking the expectation of the GARCH equation:

$$ E(V_t) = E(c + \alpha\epsilon^2_t+\beta V_{t-1}) $$ $$ V = c + \alpha V + \beta V $$ $$ V = c/(1-\alpha - \beta) $$

But now I would like to assess how many months would it take to a regular shock (e.g., one standard deviation?) to dissipate or the conditional variance converge into its unconditional variance. How would someone do this?


1 Answer 1


Due to the nature of the autoregressive components in GARCH, the shocks dissipate at an exponential rate and never fully.* So to answer your question, for a shock to dissipate fully would take an infinitely long time, and the long-run variance would never be reached, only approched however closely.

Therefore, people study half-life instead: $$ \ell:=\frac{\ln(0.5)}{\ln(\sum_{i=1}^s \alpha_i+\sum_{\beta_j}^r \beta_j)} $$ where $\alpha$s and $\beta$s are the GARCH model coefficients. In the case of a GARCH(1,1), $$ \ell=\frac{\ln(0.5)}{\ln(\alpha_1+\beta_1)}. $$ Half-life tells you after how many periods half of the shock has dissipated.

In your case, $\ell=\frac{\ln(0.5)}{\ln(0.07+0.70)}\approx 2.65$.

*This is in contrast to an ARCH(s) model where shocks completely dissipate after $s$ periods.

  • $\begingroup$ Thank you! Could you share a reference? $\endgroup$
    – John Doe
    Jun 24, 2022 at 12:23
  • 1
    $\begingroup$ @JohnDoe, I spent a good while looking for it, but I cannot find the original reference anymore. Here are some papers that employ half-life, though. (I have also used half-life in one of my empirical papers, but I do not have a reference for it in the paper, so I will not cite my paper here.) $\endgroup$ Jun 24, 2022 at 17:45

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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