21  Conditional variance processes

# required for figures  
library(ggplot2)
library(gridExtra)
# required for render latex 
library(backports)
library(latex2exp)

A stochastic process for the conditional variance can be used to specify a model of the form $$Y_t=\mathbb{E}\{Y_t\mid {\mathcal{F}}_{t-1}\}+\sqrt{\mathbb{V}\{Y_t\mid {\mathcal{F}}_{t-1}\}}\cdot u_t\text{,}$$ where the conditional mean and variance can be both stochastic. In the following section, we will explore model with constant conditional mean, i.e. $$\mathbb{E}\{Y_t\mid {\mathcal{F}}_{t-1}\}=\mu \text{,}$$ however the above expression can be generalized using AR or ARMA processes.

21.1 ARCH(p) process

The auto regressive and conditionally heteroskedastic process (ARCH) were introduced in 1982 by Robert Engle (Engle (1982)) to model the conditional variance of a time series. It is often an empirical fact in economics that the larger values of time series the larger the variance. Let’s define an ARCH process of order $p$ as: $$\begin{array}{}\text{(21.1)}& \begin{array}{rl}& Y_t=\mu +e_t\\ & e_t={\sigma }_t^2{u}_t\\ & {\sigma }_t^2=\omega +\sum _{i=1}^p{\alpha }_i{e}_{t-i}^2\end{array}\end{array}$$ where $u_t\sim \text{MDS}(0,1)$.

21.1.1 Moments

Proposition 21.1 The conditional mean of an ARCH(p) process (Equation 21.1) is equal to $$\mathbb{E}\{Y_t\mid {\mathcal{F}}_{t-1}\}=\mu \text{.}$$

Proof: Proposition 21.1

Proof. Taking the conditional expectation on the process in Equation 21.1 one obtain: $$\begin{array}{rl}\mathbb{E}\{Y_t\mid {\mathcal{F}}_{t-1}\}& =\mu +\mathbb{E}\{e_t\mid {\mathcal{F}}_{t-1}\}=\\ & =\mu +\mathbb{E}\{{\sigma }_t{u}_t\mid {\mathcal{F}}_{t-1}\}=\\ & =\mu +{\sigma }_t\mathbb{E}\{u_t\mid {\mathcal{F}}_{t-1}\}=\\ & =\mu \end{array}$$ since ${\sigma }_t$ is known and not stochastic given the information at $t-1$ in ${\mathcal{F}}_{t-1}$ and $u_t$ conditionally to ${\mathcal{F}}_{t-1}$ is a Martingale Difference Sequence with mean zero (Equation 18.3).

Proposition 21.2 The conditional variance of an ARCH(p) process (Equation 21.1) is not stochastic given the information at time $t-1$ and for a general $h\ge 1$ reads: $$\mathbb{V}\{Y_{t+h}\mid {\mathcal{F}}_t\}=\mathbb{E}\{{\sigma }_{t+h}^2\mid {\mathcal{F}}_t\}\text{.}$$

Proof: Proposition 21.2

Proof. The conditional variance of $Y$ at time $t+h$ depends on the conditional expectation of ${\sigma }_t^2$, i.e. $$\begin{array}{rl}\mathbb{V}\{Y_{t+h}\mid {\mathcal{F}}_t\}& =\mathbb{V}\{{\sigma }_{t+h}{u}_{t+h}\mid {\mathcal{F}}_t\}=\\ & =\mathbb{E}\{{\sigma }_{t+h}^2\mid {\mathcal{F}}_t\}\mathbb{E}\{u_{t+h}^2\mid {\mathcal{F}}_t\}-\mathbb{E}\{{\sigma }_{t+h}\mid {\mathcal{F}}_t{\}}^2\mathbb{E}\{u_{t+h}\mid {\mathcal{F}}_t{\}}^2=\\ & =\mathbb{E}\{{\sigma }_{t+h}^2\mid {\mathcal{F}}_t\}\mathbb{E}\{u_{t+h}^2\mid {\mathcal{F}}_t\}=\\ & =\mathbb{E}\{{\sigma }_{t+h}^2\mid {\mathcal{F}}_t\}\end{array}$$ since $\mathbb{E}\{u_{t+h}\mid {\mathcal{F}}_t\}=0$ and $\mathbb{E}\{u_{t+h}^2\mid {\mathcal{F}}_t\}=1$.

Long-term variance ARCH(p)

The long-term (unconditional) expectation of the ARCH(p) variance ${\sigma }_t^2$ reads $$\mathbb{E}\{{\sigma }_t^2\}=\frac{\omega }{1-\sum _{i=1}^p{\alpha }_i}\text{.}$$ where $\omega \ne 0$. It is clear that to ensure that the process is stationary the following condition on the ARCH parameters must be satisfied for all $i\in \{1,\dots ,p\}$, i.e. $$\sum _{i=1}^p{\alpha }_i<1\text{,}{\alpha }_i>0\text{.}$$

Example: ARCH(1) process

Example 21.1 Let’s simulate an ARCH(1) process with normal residuals, i.e.  $$\begin{array}{rl}& Y_t=\mu +e_t\\ & e_t={\sigma }_t{u}_t\\ & {\sigma }_t^2=\omega +{\alpha }_1{e}_{t-1}^2\end{array}$$ where $u_t\sim \mathcal{N}(0,1)$.

# Random seed 
set.seed(1)
# *************************************************
#                      Inputs 
# *************************************************
# Number of simulations 
t_bar <- 300 
# Long term mean of Yt 
mu <- 0.5 
# Intercept sigma2
omega <- 0.4
# ARCH parameters
alpha <- c(alpha1 = 0.25)
# *************************************************
#                    Simulation 
# *************************************************
# Long-term variance 
e_sigma2 <- omega / (1 - sum(alpha))
# Starting point
Yt <- rep(mu, 2)
# Initialization 
sigma2_t <- rep(e_sigma2, 4)
# Simulate standardized residuals 
u_t <- rnorm(t_bar, 0, 1) 
# Initialize residuals 
e_t <- sigma2_t * u_t
for(t in 2:t_bar){
  # ARCH components
  sum_arch <- alpha[1] * e_t[t-1]^2
  # Next-step variance 
  sigma2_t[t] <- omega + sum_arch
  # Update simulated residuals 
  e_t[t] <- sqrt(sigma2_t[t]) * u_t[t]
  # Update time series 
  Yt[t] <- mu + e_t[t]
}

Figure 21.1: On the top a ARCH(1) simulation with its long term mean (red). On the bottom the correspondent stochastic variance with its long term mean (blue).

Example: ARCH(3) process

Example 21.2 Let’s simulate an ARCH(3) process with normal residuals, i.e.  $$\begin{array}{rl}& Y_t=\mu +e_t\\ & e_t={\sigma }_t{u}_t\\ & {\sigma }_t^2=\omega +{\alpha }_1{e}_{t-1}^2+{\alpha }_2{e}_{t-2}^2+{\alpha }_3{e}_{t-3}^3\end{array}$$ where $u_t\sim \mathcal{N}(0,1)$.

# Random seed 
set.seed(2)
# *************************************************
#                      Inputs 
# *************************************************
# Number of simulations 
t_bar <- 300 
# Long term mean of Yt 
mu <- 0.5 
# Intercept sigma2
omega <- 0.4
# ARCH parameters
alpha <- c(alpha1 = 0.25, alpha2 = 0.05, alpha3 = 0.02)
# *************************************************
#                    Simulation 
# *************************************************
# Long-term variance 
e_sigma2 <- omega / (1 - sum(alpha))
# Starting point
Yt <- rep(mu, 4)
# Initialization 
sigma2_t <- rep(e_sigma2, 4)
# Simulate standardized residuals 
u_t <- rnorm(t_bar, 0, 1) 
# Initialize residuals 
e_t <- sigma2_t * u_t
for(t in 4:t_bar){
  # ARCH components
  sum_arch <- alpha[1] * e_t[t-1]^2 + alpha[2] * e_t[t-2]^2 + alpha[3] * e_t[t-3]^2
  # Next-step variance 
  sigma2_t[t] <- omega + sum_arch
  # Update simulated residuals 
  e_t[t] <- sqrt(sigma2_t[t]) * u_t[t]
  # Update time series 
  Yt[t] <- mu + e_t[t]
}

Figure 21.2: On the top a ARCH(3) simulation with its long term mean (red). On the bottom the correspondent stochastic variance with its long term mean (blue).

21.2 GARCH(p,q) process

As done with the ARCH(p), with generalized auto regressive conditional heteroskedasticity (GARCH) we model the dependency of the conditional second moment. It represents a more parsimonious way to express the conditional variance. A GARCH(p,q) processx is defined as: $$\begin{array}{rl}& Y_t=\mu +e_t\\ & e_t={\sigma }_t{u}_t\\ & {\sigma }_t^2=\omega +\sum _{i=1}^p{\alpha }_i{e}_{t-i}^2+\sum _{j=1}^q{\beta }_j{\sigma }_{t-j}^2\end{array}$$ where $u_t\sim \text{MDS}(0,1)$.

Long-term variance GARCH(p,q)

Regarding the variance, the long-term expectation of ${\sigma }_t^2$ reads $$\begin{array}{}\text{(21.2)}& \mathbb{E}\{{\sigma }_t^2\}=\frac{\omega }{1-\sum _{i=1}^p{\alpha }_i-\sum _{j=1}^q{\beta }_j}\text{,}\end{array}$$ where $\omega \ne 0$. It is clear that to ensure that the process is stationary the following condition on the GARCH parameters must be satisfied: $$\sum _{i=1}^p{\alpha }_i+\sum _{j=1}^q{\beta }_j<1\text{,}$$ with ${\alpha }_i\ge 0$ and ${\beta }_j\ge 0$ for all $i\in \{1,\dots ,p\}$ and $j\in \{1,\dots ,q\}$.

Example: GARCH(1,1) process

Example 21.3 Let’s simulate an GARCH(1,1) process with normal residuals, i.e.  $$\begin{array}{rl}& Y_t=\mu +e_t\\ & e_t={\sigma }_t{u}_t\\ & {\sigma }_t^2=\omega +{\alpha }_1{e}_{t-1}^2+{\beta }_1{\sigma }_{t-1}^2\end{array}$$ where $u_t\sim \mathcal{N}(0,1)$.

# Random seed 
set.seed(3)
# *************************************************
#                      Inputs 
# *************************************************
# Number of simulations 
t_bar <- 300 
# Long term mean of Yt 
mu <- 0.5 
# Intercept sigma2
omega <- 0.4
# ARCH parameters
alpha <- c(alpha1 = 0.25)
# GARCH parameters
beta <- c(beta1 = 0.25)
# *************************************************
#                    Simulation 
# *************************************************
# Long-term variance 
e_sigma2 <- omega / (1 - sum(alpha) - sum(beta))
# Starting point
Yt <- rep(mu, 2)
# Initialization 
sigma2_t <- rep(e_sigma2, 4)
# Simulate standardized residuals 
u_t <- rnorm(t_bar, 0, 1) 
# Initialize residuals 
e_t <- sigma2_t * u_t
for(t in 2:t_bar){
  # ARCH components
  sum_arch <- alpha[1] * e_t[t-1]^2
  # GARCH components 
  sum_garch <- beta[1] * sigma2_t[t-1] 
  # Next-step variance 
  sigma2_t[t] <- omega + sum_arch + sum_garch
  # Update simulated residuals 
  e_t[t] <- sqrt(sigma2_t[t]) * u_t[t]
  # Update time series 
  Yt[t] <- mu + e_t[t]
}

Figure 21.3: On the top a GARCH(1,1) simulation with its long term mean (red). On the bottom the correspondent stochastic variance with its long term mean (blue).

Example: GARCH(2,3) process

Example 21.4 Let’s simulate a GARCH(2,3) process with normal residuals, i.e $$\begin{array}{rl}& Y_t=\mu +e_t\\ & e_t={\sigma }_t{u}_t\\ & {\sigma }_t^2=\omega +{\alpha }_1{e}_{t-1}^2+{\alpha }_2{e}_{t-2}^2+{\beta }_1{\sigma }_{t-1}^2+{\beta }_2{\sigma }_{t-2}^2+{\beta }_3{\sigma }_{t-3}^2\end{array}$$ where $u_t\sim \mathcal{N}(0,1)$.

# Random seed 
set.seed(4)
# *************************************************
#                      Inputs 
# *************************************************
# Number of simulations 
t_bar <- 300 
# Long term mean of Yt 
mu <- 0.5 
# Intercept sigma2
omega <- 0.4
# ARCH parameters
alpha <- c(alpha1 = 0.05, alpha2 = 0.03)
# GARCH parameters
beta <- c(beta1 = 0.55, beta2 = 0.15, beta3 = 0.05)
# *************************************************
#                    Simulation 
# *************************************************
# Long-term variance 
e_sigma2 <- omega / (1 - sum(alpha) - sum(beta))
# Starting point
Yt <- rep(mu, 4)
# Initialization 
sigma2_t <- rep(e_sigma2, 4)
# Simulate standardized residuals 
u_t <- rnorm(t_bar, 0, 1) 
# Initialize residuals 
e_t <- sigma2_t * u_t
for(t in 4:t_bar){
  # ARCH components
  sum_arch <- alpha[1] * e_t[t-1]^2 + alpha[2] * e_t[t-2]^2
  # GARCH components 
  sum_garch <- beta[1] * sigma2_t[t-1] + beta[2] * sigma2_t[t-2] + beta[3] * sigma2_t[t-3]
  # Next-step variance 
  sigma2_t[t] <- omega + sum_arch + sum_garch
  # Update simulated residuals 
  e_t[t] <- sqrt(sigma2_t[t]) * u_t[t]
  # Update time series 
  Yt[t] <- mu + e_t[t]
}

Figure 21.4: On the top a GARCH(2,3) simulation with its long term mean (red). On the bottom the correspondent stochastic variance with its long term mean (blue).

Example: GARCH(3,2) process

Example 21.5 Let’s simulate a GARCH(3,2) process with normal residuals, i.e.  $$\begin{array}{rl}& Y_t=\mu +e_t\\ & e_t={\sigma }_t{u}_t\\ & {\sigma }_t^2=\omega +{\alpha }_1{e}_{t-1}^2+{\alpha }_2{e}_{t-2}^2+{\alpha }_3{e}_{t-3}^2+{\beta }_1{\sigma }_{t-1}^2+{\beta }_2{\sigma }_{t-2}^2\end{array}$$ where $u_t\sim \mathcal{N}(0,1)$.

# Random seed 
set.seed(5)
# *************************************************
#                      Inputs 
# *************************************************
# Number of simulations 
t_bar <- 300 
# Long term mean of Yt 
mu <- 0.5 
# Intercept sigma2
omega <- 0.1
# ARCH parameters
alpha <- c(alpha1 = 0.35, alpha2 = 0.15, alpha3 = 0.05)
# GARCH parameters
beta <- c(beta1 = 0.25, beta2 = 0.03)
# *************************************************
#                    Simulation 
# *************************************************
# Long-term variance 
e_sigma2 <- omega / (1 - sum(alpha) - sum(beta))
# Starting point
Yt <- rep(mu, 4)
# Initialization 
sigma2_t <- rep(e_sigma2, 4)
# Simulate standardized residuals 
u_t <- rnorm(t_bar, 0, 1) 
# Initialize residuals 
e_t <- sigma2_t * u_t
for(t in 4:t_bar){
  # ARCH components
  sum_arch <- alpha[1] * e_t[t-1]^2 + alpha[2] * e_t[t-2]^2 + alpha[3] * e_t[t-3]^2
  # GARCH components 
  sum_garch <- beta[1] * sigma2_t[t-1] + beta[2] * sigma2_t[t-2]
  # Next-step variance 
  sigma2_t[t] <- omega + sum_arch + sum_garch
  # Update simulated residuals 
  e_t[t] <- sqrt(sigma2_t[t]) * u_t[t]
  # Update time series 
  Yt[t] <- mu + e_t[t]
}

Figure 21.5: On the top a GARCH(3,2) simulation with its long term mean (red). On the bottom the correspondent stochastic variance with its long term mean (blue).

21.3 IGARCH

Many variants of the standard GARCH process were developed in literature. For example, the Integrated Generalized Auto regressive Conditional heteroskedasticity (IGARCH(p,q)) is a restricted version of the GARCH, where the persistent parameters sum up to one, and imply a unit root in the GARCH process with the condition $$\sum _{i=1}^p{\alpha }_i+\sum _{j=1}^q{\beta }_j=1\text{.}$$ In GARCH(p, q), the sum of coefficients being less than 1 ensures stationarity (finite unconditional variance) while in IGARCH(p, q), the sum is exactly 1, so the process is nonstationary in variance: the conditional variance has a persistent memory, and the shocks to volatility accumulate over time. The process is strictly stationary under some conditions (see Nelson (1990)), but it has infinite unconditional variance.

Example: iGARCH(1,1) process

Example 21.6 Let’s simulate an iGARCH(1,1) process with normal residuals, i.e.  $$\begin{array}{rl}& Y_t=\mu +e_t\\ & e_t={\sigma }_t{u}_t\\ & {\sigma }_t^2=\omega +{\alpha }_1{e}_{t-1}^2+(1-{\alpha }_1){\sigma }_{t-1}^2\end{array}$$ where $u_t\sim \mathcal{N}(0,1)$.

# Random seed 
set.seed(6)
# *************************************************
#                      Inputs 
# *************************************************
# Number of simulations 
t_bar <- 300 
# Long term mean of Yt 
mu <- 0.5 
# Intercept sigma2
omega <- 0.4
# ARCH parameters
alpha <- c(alpha1 = 0.35)
# *************************************************
#                    Simulation 
# *************************************************
# Starting point
Yt <- rep(mu, 2)
# Initialization 
sigma2_t <- rep(1, 2)
# Simulate standardized residuals 
u_t <- rnorm(t_bar, 0, 1) 
# Initialize residuals 
e_t <- sigma2_t * u_t
for(t in 2:t_bar){
  # ARCH components
  sum_arch <- alpha[1] * e_t[t-1]^2
  # GARCH components 
  sum_garch <- (1 - alpha[1]) * sigma2_t[t-1] 
  # Next-step variance 
  sigma2_t[t] <- omega + sum_arch + sum_garch
  # Update simulated residuals 
  e_t[t] <- sqrt(sigma2_t[t]) * u_t[t]
  # Update time series 
  Yt[t] <- mu + e_t[t]
}

Figure 21.6: On the top a GARCH(1,1) simulation with its long term mean (red). On the bottom the correspondent stochastic variance with its long term mean (blue).

21.4 GARCH-M

The GARCH in-mean (GARCH-M) model adds a stochastic term into the mean equation. This is motivated especially in financial theories (e.g., risk-return trade-off) suggesting that expected returns may depend on volatility, i.e. $$\begin{array}{rl}& Y_t=\mu +\lambda {\sigma }_t^2+e_t\\ & e_t={\sigma }_t{u}_t\\ & {\sigma }_t^2\sim \text{GARCH}(p,q)\end{array}$$ The effect of the parameter $\lambda$ is a shift of the mean of the process. If $Y_t$ are for example financial returns a $\lambda >0$ would imply that higher volatility increases expected return, consistent with risk-premium theories. Instead, when $\lambda <0$, it could reflect behavioral phenomena or model misspecification. The unconditional mean of $Y_t$ became $$\mathbb{E}\{Y_t\}=\mu +\lambda \mathbb{E}\{{\sigma }_t^2\}\text{,}$$ while the conditional mean $$\mathbb{E}\{Y_{t+h}\mid {\mathcal{F}}_t\}=\mu +\lambda \mathbb{E}\{{\sigma }_{t+h}^2\mid {\mathcal{F}}_t\}\text{.}$$

Example: GARCH-M(1,1) process

Example 21.7 Let’s simulate an GARCH-M(1,1) process with normal residuals, i.e.  $$\begin{array}{rl}& Y_t=\mu +\lambda {\sigma }_t^2+e_t\\ & e_t={\sigma }_t{u}_t\\ & {\sigma }_t^2=\omega +{\alpha }_1{e}_{t-1}^2+{\beta }_1{\sigma }_{t-1}^2\end{array}$$ where $u_t\sim \mathcal{N}(0,1)$.

# Random seed 
set.seed(7)
# *************************************************
#                      Inputs 
# *************************************************
# Number of simulations 
t_bar <- 300 
# Long term mean of Yt 
mu <- 0.5 
# Intercept sigma2
omega <- 0.4
# ARCH parameters
alpha <- c(alpha1 = 0.15)
# Mean param 
lambda <- 2.5
# GARCH parameters
beta <- c(beta1 = 0.35)
# *************************************************
#                    Simulation 
# *************************************************
# Long-term variance 
e_sigma2 <- omega / (1 - sum(alpha) - sum(beta))
# Starting point
Yt <- rep(mu, 2)
# Initialization 
sigma2_t <- rep(e_sigma2, 4)
# Simulate standardized residuals 
u_t <- rnorm(t_bar, 0, 1) 
# Initialize residuals 
e_t <- sigma2_t * u_t
for(t in 2:t_bar){
  # ARCH components
  sum_arch <- alpha[1] * e_t[t-1]^2
  # GARCH components 
  sum_garch <- beta[1] * sigma2_t[t-1] 
  # Next-step variance 
  sigma2_t[t] <- omega + sum_arch + sum_garch
  # Update simulated residuals 
  e_t[t] <- sqrt(sigma2_t[t]) * u_t[t]
  # Update time series 
  Yt[t] <- mu + sigma2_t[t] * lambda + e_t[t]
}

Figure 21.7: On the top a GARCH-M(1,1) simulation with its long term mean (red). On the bottom the correspondent stochastic variance with its long term mean (blue).

Engle, Robert F. 1982. “Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation.” Econometrica 50 (4): 987–1007. https://doi.org/10.2307/1912773.

Nelson, Daniel B. 1990. “Stationarity and Persistence in the GARCH(1,1) Model.” Econometric Theory 6 (3): 318–34. http://www.jstor.org/stable/3532198.