GARCH(1,1) Conditional and Unconditional Variance
The GARCH(1,1) model specifies daily returns as $r_t = \sigma_t \varepsilon_t$ with $\varepsilon_t \sim N(0,1)$ i.i.d., and conditional variance:
$\sigma_t^2 = \omega + \alpha\, r_{t-1}^2 + \beta\, \sigma_{t-1}^2$
You are given parameters $\omega = 0.000002$, $\alpha = 0.08$, $\beta = 0.91$, yesterday's return $r_{t-1} = -0.03$, and yesterday's conditional variance $\sigma_{t-1}^2 = 0.0004$.
- Compute today's conditional variance $\sigma_t^2$ and daily volatility $\sigma_t$.
- Compute the unconditional (long-run) variance and annualized volatility.
- What is the persistence $\alpha + \beta$, and what does it imply about the half-life of variance shocks?
Hints
- Just plug the given values directly into the GARCH(1,1) recursion $\sigma_t^2 = \omega + \alpha r_{t-1}^2 + \beta \sigma_{t-1}^2$.
- For the unconditional variance, set $\sigma^2 = \omega + (\alpha + \beta)\sigma^2$ and solve. The stationarity condition $\alpha + \beta < 1$ must hold.
- The half-life of a variance shock under GARCH is $\ln 2 / (-\ln(\alpha + \beta))$. With $\alpha + \beta = 0.99$, shocks are extremely persistent.
Worked Solution
How to Think About It: GARCH(1,1) is the workhorse volatility model in finance. It captures two stylized facts: volatility clusters (big moves follow big moves) and mean-reverts (vol eventually returns to a long-run level). The conditional variance equation is just an exponentially weighted moving average of past squared returns, with a pull toward the long-run variance. When you see a large return (like $-3\%$ here), the model ratchets up tomorrow's vol forecast. How much it ratchets up depends on $\alpha$ (reaction to new shocks) and how quickly it fades depends on $\beta$ (persistence of old variance).
Quick Estimate: Yesterday's vol was $\sqrt{0.0004} = 2\%$ and we saw a $-3\%$ move -- that's a
Formal Solution:
Part 1: Today's conditional variance
$\sigma_t^2 = \omega + \alpha\, r_{t-1}^2 + \beta\, \sigma_{t-1}^2$
$= 0.000002 + 0.08 \times (-0.03)^2 + 0.91 \times 0.0004$
$= 0.000002 + 0.08 \times 0.0009 + 0.91 \times 0.0004$
$= 0.000002 + 0.000072 + 0.000364 = 0.000438$
Daily volatility: $\sigma_t = \sqrt{0.000438} \approx 0.02093 = 2.09\%$
Yesterday's vol was $\sqrt{0.0004} = 2.00\%$, so today's forecast is higher due to the large negative return.
Part 2: Unconditional (long-run) variance
For stationarity, require $\alpha + \beta < 1$. Here $0.08 + 0.91 = 0.99 < 1$. The unconditional variance satisfies $\bar{\sigma}^2 = \omega + \alpha \bar{\sigma}^2 + \beta \bar{\sigma}^2$, giving:
$\bar{\sigma}^2 = \frac{\omega}{1 - \alpha - \beta} = \frac{0.000002}{1 - 0.99} = \frac{0.000002}{0.01} = 0.0002$
Long-run daily vol: $\bar{\sigma} = \sqrt{0.0002} \approx 1.414\%$
Annualized vol: $\bar{\sigma}_{\text{ann}} = 1.414\% \times \sqrt{252} \approx 22.4\%$
Note: today's conditional variance $0.000438$ is well above the long-run level $0.0002$. The model predicts vol will gradually decay back toward
Part 3: Persistence and half-life
Persistence: $\alpha + \beta = 0.99$
This is very high persistence -- shocks to variance decay extremely slowly. The half-life of a variance shock is:
$t_{1/2} = \frac{\ln 2}{-\ln(\alpha + \beta)} = \frac{\ln 2}{-\ln(0.99)} \approx \frac{0.693}{0.01005} \approx 69 \text{ days}$
This means if today's variance spikes, it takes about 69 trading days (roughly 3 months) for the excess variance to decay by half.
Practical Interpretation: - The decomposition of variance into components: $\omega = 0.000002$ (constant floor, tiny), $\alpha r_{t-1}^2 = 0.000072$ (reaction to yesterday's shock, 16% of today's variance), $\beta \sigma_{t-1}^2 = 0.000364$ (inherited from past, 83% of today's variance). - The model is heavily persistence-driven ($\beta = 0.91$), with moderate shock reactivity ($\alpha = 0.08$). This is typical for equity indices. - With $\alpha + \beta = 0.99$ close to 1, this model is nearly IGARCH -- variance shocks are almost permanent.
Answer:
- $\sigma_t^2 = 0.000438$, daily vol $\approx 2.09\%$
- $\bar{\sigma}^2 = 0.0002$, long-run daily vol $\approx 1.41\%$, annualized $\approx 22.4\%$
- Persistence $= 0.99$, half-life $\approx 69$ trading days
Intuition
GARCH(1,1) is the minimum viable volatility model for most financial applications. The conditional variance equation is really just a weighted average: $\omega/(1-\alpha-\beta)$ pulls you toward the long-run variance, $\alpha$ controls how much yesterday's surprise matters, and $\beta$ controls inertia. A high $\alpha + \beta$ (close to 1) means the model has a long memory -- volatility regimes persist for months. This particular parameterization ($\alpha + \beta = 0.99$) is very typical for equity index returns, where vol clusters are well-documented.
The key practical takeaway: the unconditional variance gives you the long-run risk level, but the conditional variance is what matters for short-term risk management. Today's conditional vol is