2.1 AR(1)

\[\begin{align*} \tilde{Z}_t &= \phi_1 \tilde{Z}_{t-1} + a_t\\ Z_{n+1} &= \mu + \phi_1 (Z_{n} - \mu ) + a_{n+1} \end{align*}\] Observe \(z_1, ..., z_n\). Our forecast is \[\begin{align*} \hat{z}_{n}(1) &= \mathbb{E}_n [ Z_{n+1} ]\\ & = \mu + \phi_1 (\mathbb{E}_n [Z_n] - \mu ) + \mathbb{E}_n[a_{n+1}]\\ & = \mu + \phi_1 (z_n - \mu ) + 0 \end{align*}\]

By assumption \(\mathbb{E}_n[a_{n+1}]=0\), because \(a_{n+1}\) takes place after \(z_n, z_{n-1},...\).

\[\begin{align*} \hat{z}_{n}(1) &= \mu + \phi_1 (z_n - \mu ) \end{align*}\]

Since \(\mu\) and \(\phi_1\) are unknown, so to forecast, we replace parameters by estimates, \(\hat{\mu}\) and \(\hat{\phi}_1\) and our final forecast is

\[\begin{align*} \hat{z}_{n}(1) &= \hat{\mu} + \hat{\phi}_1 (z_n - \hat{\mu} ) \end{align*}\]

We can see that \[\begin{align*} \hat{z}_{n}(2) &= \mathbb{E}_n [ Z_{n+2} ]\\ & = \mu + \phi_1 (\mathbb{E}_n [Z_{n+1}] - \mu ) + \mathbb{E}_n[a_{n+2}]\\ & = \mu + \phi_1 (\hat{z}_{n}(1) - \mu )\\ &= \mu + \phi_1^2 (z_n - \mu ) \end{align*}\]

In general,

\[\begin{align*} \hat{z}_{n}(l) &= \mu + \phi_1^l (z_n - \mu ) \end{align*}\]

for \(l=1,2,...\)

Note that for stationary AR(1), \(\phi_1^l \to 0\) when \(l \to \infty\) because \(|\phi_1|<1\).

2.2 MA(1)

\[ \tilde{Z}_t = a_t - \theta_1 a_{t-1} \]

Case 1: \(l>1\)

\[ Z_{n+l} = \mu + a_{n+l} - \theta_1 a_{n+l-1}\\ \hat{z}_{n}(l) = \mathbb{E}_n(Z_{n+l}) = \mu + 0 - 0 = \mu \]

Case 2: \(l=1\)

Since \(Z_{n+1} = \mu + a_{n+1} - \theta_1 a_{n}\), we have \[ \hat{z}_{n}(1) = \mathbb{E}_n(Z_{n+1}) = \mu + 0 - \theta_1 \mathbb{E}_n(a_{n}) \]

Since \(a_n\) depends on current and past observations \(z_t\)’s, \(\mathbb{E}_n(a_n)\) is not 0 any more. Thus we estimate \(\mathbb{E}_n(a_n)\) by \(\hat{a}_n = z_n - \hat{z}_n =\) residual at time \(n\).

Therefore,

\[ \hat{z}_{n}(1) = \mu + 0 - \theta_1 \hat{a}_n \] To get a actual forecast, we plug in \(\hat{\mu}\) and \(\hat{\theta}_1\).

2.3 General Forecast

A general forecast formula of ARIMA

Suppose \(Z_t \sim ARIMA(p,d,q)\), which can be considered as \(Z_t \sim ARMA(p+d,q)\) (nonstationary)

\[\begin{align*} Z_t &= C + \Phi_1 Z_{t-1} + ... + \Phi_{p+d} Z_{t-p-d} + a_t - \theta_1 a_{t-1} ... - \theta_q a_{t-q} \end{align*}\]

\[\begin{align*} Z_{n+l} &= C + \Phi_1 Z_{n+l-1} + ... + \Phi_{p+d} Z_{n+l-p-d} + a_t - \theta_1 a_{n+l-1} ... - \theta_q a_{n+l-q} \end{align*}\]

The general formula \[\begin{align*} \hat{z}_{n}(l) =& \mathbb{E}[Z_{n+l}|\mathcal{F}_n] \\ =& C + \Phi_1 \mathbb{E}[Z_{n+l-1}|\mathcal{F}_n] + ... + \Phi_{p+d} \mathbb{E}[Z_{n+l-p-d}|\mathcal{F}_n] + \\ &\mathbb{E}[a_{n+l}|\mathcal{F}_n] - \theta_1 \mathbb{E}[a_{n+l-1}|\mathcal{F}_n] - ... - \theta_q \mathbb{E}[a_{n+l-q}|\mathcal{F}_n] \end{align*}\]

\[\begin{align*} \mathbb{E}[Z_{s}|\mathcal{F}_n]=\begin{cases} z_s, \quad \quad \text{ if } s \leq n\\ \hat{z}_n(s-n), \quad \text{ if } s > n \end{cases} \end{align*}\]

\[\begin{align*} \mathbb{E}[a_{s}|\mathcal{F}_n]=\begin{cases} \hat{a}_s=z_s - \hat{z}_{s-1}(1), \quad \text{ if } s \leq n\\ 0, \quad \quad \text{ if } s >n \end{cases} \end{align*}\]

Define

\[\begin{align*} \hat{Z}_{n+l}=\begin{cases} \hat{z}_n(l), \quad \quad \text{ if } l \geq 1\\ z_{n+l}, \quad \text{ if } l < 0 \end{cases} \end{align*}\]

\[\begin{align*} [a_{s}]=\begin{cases} \hat{a}_s, \quad \text{ if } s \leq n\\ 0, \quad \quad \text{ if } s > n \end{cases} \end{align*}\]

Then \[ \hat{z}_n(l) = C + \phi_1 \hat{Z}_{n+l-1} + ... + \phi_{p+d} \hat{Z}_{n+l-p-d} - \theta_1 [a_{n+l-1}] - ... - \theta_q [a_{n+l-q}] \]

3.1 Long Run Mean of Forecast from Stationary Time Series Models

Let’s use AR(1) as an example.

\[\begin{align*} \hat{z}_n(l)&=\mathbb{E}_n[Z_{n+l}]\\ &=\mathbb{E}_n[u + \phi_1 (Z_{n+l-1}-u) + a_{n+l}]\\ &=\mathbb{E}_n[u + \phi_1^2 (Z_{n+l-2}-u) + \phi_1 a_{n+l-1} + a_{n+l}]\\ & ...\\ &=\mathbb{E}_n[\mu + a_{n+l} + \phi_1 a_{n+l-1} + \phi_1^2 a_{n+l-2} ... ]\\ &=\mu + 0 ... + 0 + \phi_1^l a_n + \phi_1^{(l+1)} a_{n-1} ... \end{align*}\] Note that \(|\phi_1|<1\) for stationary time series model. Therefore, \(\phi_1^l \to 0\) as \(l \to \infty\). Therefore, the limit of the forecast is

\[\begin{align*} \lim_{l \to \infty} \hat{z}_n(l) = \mu + \lim_{l \to \infty} (\phi_1^l \hat{a}_n + \phi_1^{(l+1)} \hat{a}_{n-1} ...) = \mu \end{align*}\]

Therefore, the forecast converges to the mean for the stationary time series model. This result holds for AR(1), and it is generally true for all stationary AR, MA, ARMA models.

3.2 Long Run Mean of Forecast from Nonstationary Time Series Models

The forecast for nonstationary time series model do no revert to a fixed mean.

Consider a random walk \(Z_t=Z_{t-1}+a_t\), we have

\[\begin{align*} \hat{z}_n(l)&= \begin{cases} z_n, \quad l=1\\ \hat{z}_n(l-1), \quad \quad l > 1 \end{cases} \end{align*}\]

So, we combine the cases and have

\[ \hat{z}_n(l) = z_n \text{ for }l \geq 1 \]

Thus, different realization of the same random walk model will result in forecasts converge to different values.

On the other hand, different realizations of the same stationary time series model will always produce the forecasts which converge to the same constant, \(\mu\).

3.3 Long Run Variance of Forecast for Stationary and Nonstationary Models

Next, consider the limit of the variance of the forecast error of a stationary time series model. Let’s take AR(1) as an example.

\[ \lim_{l \to \infty} \text{Var}(Z_{n+l} - \hat{z}_{n}(l))\\ =\lim_{l \to \infty} \sigma_a^2(1+\phi_1^2+(\phi_1^2)^2+(\phi_1^{l-1})^2+...)\\ =\sigma_a^2 \sum_{l=0}^{\infty} (\phi_1^2)^l \]

Thus the prediction “bands”about the forecast profile become parallel lines for large lead time (\(l \to \infty\)).

In fact, for AR(1), we can show

\[ \text{Var}(Z_{n+l} - \hat{z}_{n}(l)) = \sigma_a^2(1 + \phi_1^2 + \phi_1^4+\phi_1^{2(l-1)}) \]

On the other hand, consider the limit of the variance of the forecast error for a nonstationary time series, for example, random walk.

\[ (1-B)Z_t = a_t\\ Z_t = (1-B)^{-1} a_t\\ Z_t = (1+B+B^2+B^3...) a_t\\ Z_t = a_t + a_{t-1} + a_{t-2} + ...\\ \] Therefore, the variance becomes

\[ \text{Var}(Z_{n+l} - \hat{z}_n(l)) = \sigma^2_a \sum_{j=0}^{l-1} 1 = \sigma^2_a * l \] So the variance \(\text{Var}(Z_{n+l} - \hat{z}_n(l))\) tends to infinity when \(l \to \infty\), so the prediction band expands.