# Predictive density and likelihood evaluation at time t+1 of GARCH model

I am new to forecast and I would appreciate any help.

I want to do Bayesian estimation of GARCH models. I read a similar question here, but I have some additional questions. The model is

$$y_i=\sigma_i\epsilon_i,$$ $$\sigma^2_i = \alpha_0 + \alpha_1 y^2_{i-1} + \alpha_2 \sigma^2_{i-1},$$ $$\text{where } \epsilon_i \overset{iid}{\sim}\mathcal{N}(0,1),\ i=1,2,\cdots,T.$$

The parameters of the model are estimated using the first $t$ observations, and these estimates are used to make one-step-ahead forecasts for the remaining $T-t$ periods. The vignette of the package stochvol says (page 23, Algortihm 1) to do the following steps:

Algorithm 1 (Predictive density and likelihood evaluation at time t + 1)

1. Obtain M posterior draws of $\theta, \text{where } \theta=(\alpha_0,\ \alpha_1,\ \alpha_2).$
2. Obtain M draws from the conditional distribution $\sigma_{{t+1}|[1:t]}|y_{[1:t]},\theta_{[1:t]}$ by computing $\sigma^{(m)}_{{t+1}|[1:t]} =\sqrt{ \alpha_{0,[1:t]}^{(m)} + \alpha_{1,[1:t]}^{(m)} (y_t^{o})^2+\alpha_{2,[1:t]}^{(m)} ( \sigma^{(m)}_{t,[1:t]})^2 }$
3. To obtain $\text{PL}_{t+1}$, average over M densities of normal distribution with mean $(1,y_t) \times \beta_{[1:t]}^{(m)}$ and standard deviation $\exp( \sigma_{t+1,[1:t]}^{(m)})$, each evaluated at $y_{t+1}^{o}$ for $m=1,2,\cdots, M.$
4. To obtain M draws from the predictive distribution, simulate from a normal distribution with mean $(1,y_t) \times \beta_{[1:t]}^{(m)}$ and Standard deviation $\exp( \sigma_{t+1,[1:t]}^{(m)})$ for $m=1,2,\cdots, M.$,

where by using the superscript $(^o)$ in $y_{[1:t]}$, we follow Geweke and Amisano (2010) and denote ex post realizations (observations) for the set of points in time ${1,2,\cdots,t}$ of the ex ante random values $y_{[1:t]}$ (page 22).

My questions are:

1. For step 2: the term $(y_t^{o})^2$ represents the observed price at time $t$? I read here (page 7) that starting from the GARCH(1,1) equation for $\sigma_{t}^2$ , we can derive our forecast for next period's variance, $\hat{\sigma}_{t+1}^2$ by

$$\hat{\sigma}_{t+1}^2 =\alpha_0 + \alpha_1 E(y_t^2|\mathcal{F}_{t-1}) + \alpha_2 \sigma_t^{2} = \alpha_0 +(\alpha_1 +\alpha_2)\sigma_t^{2}$$ While we use the observed value at step 2, we use the term $E(y_t^2|\mathcal{F}_{t-1})$ at the last equation. Please, can you explain this to me. Moreover, why do we use the term $E(y_t^2|\mathcal{F}_{t-1})$ and not $E(y_t^2|\mathcal{F}_{t})$?

1. For step 3: My question is similar. The terms $(y_{t+1}^{o})^2$ are the true values? Here, I don't have $\beta$-parameters, so the mean is zero?
2. Could you please propose a book that will help me?
• stochvol does not handle Bayesian GARCH estimation but stochastic volatility. The vignette just shows a comparison between stochastic volatility and a GARCH method, however, a random walk MCMC is implemented which is (to my knowledge) not included in the stochvol package per se. Instead, use for example the bayesGARCH package! – muffin1974 Feb 19 '16 at 17:09
• Just a brief notice regarding your question 2: $\beta$ is any way to compute the expected mean, this is not covered by GARCH estimation as this does only focus on volatilities. $\beta$ would come up if you additionally impose an ARMA process (or whatever idea you have regarding the dynamics of the mean), otherwise just set it to 0 – muffin1974 Feb 19 '16 at 17:14
• Thank you for your answer. I edit my question. I just want to understand the steps of the algorithm and to write my code. – F.F. Feb 19 '16 at 17:24

1. Question: The identity $E(y_t ^2|F_{t-1})=(y_t^0) ^2$ holds, because in this setting you assume that the mean dynamics of $y_t$ is flat, which means the best predictor of $y_t$ at time $t-1$ is just $y_{t-1}$. So you are correct, $y_t^0$ is the $t-1$ observed return. $E(y_t ^2|F_{t})$ cannot be used because given you could now this value there would be no need to predict anything as $E(y_t ^2|F_{t})=y_t$. However, the best predictor in terms of mean-squared error is $E(y_t ^2|F_{t-1})$ thats is the reason why you use this value.
2. Yes, you are correct, and as I wrote in the comments, the underlying assumption is that of a flat mean dynamic, so $\beta=0$.
• Hi! Thank you very much. To be honest, I don't understand why to take mean. Suppose that we have data $y_1,y_2,\cdots,y_t$ and estimates $\hat{\theta}$. We know that $\sigma^2_{t+1}=\alpha_0+\alpha_1y^2_t + \alpha_2\sigma^2_t$. At time $t$, $y_{t+1}$ is not known, but $y_t$ is known. So, we predict $\hat{\sigma}^2_{t+1}=\hat{\alpha_0}+\hat{\alpha_1}y^2_t + \hat{\alpha_2}\sigma^2_t$. So, why to take mean? Maybe, I haven't understoond properly the definition of prediction. – F.F. Feb 19 '16 at 19:57
• Can you try to be more precise, I do not get what you mean with 'take mean'? Are you wondering why the formula says $E(y_t|F_t)$ although in the end you just use $y_{t-1}$? – muffin1974 Feb 20 '16 at 22:04