Q: Why is $ z_{\alpha/2} $ used, and why is it approximately 1.96 for a 95% interval?

The $ z_{\alpha/2} $ value comes from the standard normal distribution. When constructing a $ (1-\alpha) \times 100\% $ prediction interval, we want to capture the central $ (1-\alpha) $ portion of the distribution. This means $ \alpha/2 $ probability is left in each tail. For a 95% interval, $ \alpha = 0.05 $, so $ \alpha/2 = 0.025 $. The value $ z_{0.025} $ is the point such that $ P(Z > z_{0.025}) = 0.025 $ for a standard normal random variable $ Z $, which is approximately 1.96.

Q: How does the autoregressive parameter $ \phi $ influence the width of the prediction interval?

The parameter $ \phi $ determines the persistence of the series. If $ \phi $ is close to 0, the process is closer to white noise, and the interval width quickly stabilizes to $ \pm z_{\alpha/2} \sigma_{\epsilon} $. If $ |\phi| $ is large (closer to 1), the influence of past values is stronger, and the uncertainty accumulates more rapidly over time, leading to a much wider prediction interval for larger $ h $.

Q: What happens to the prediction interval as $ h $ approaches infinity for a stationary AR(1) process?

For a stationary AR(1) process (where $ |\phi| < 1 $), as $ h \to \infty $, the point forecast $ \hat{Y}_{t+h|t} $ converges to the unconditional mean of the process, $ E[Y_t] = c / (1-\phi) $. The sum $ \sum_{j=0}^{h-1} \phi^{2j} $ converges to $ 1 / (1-\phi^2) $. Thus, the forecast error variance converges to $ \sigma_{\epsilon}^2 / (1-\phi^2) $, which is the unconditional variance of the AR(1) process, $ Var(Y_t) $. The prediction interval will stabilize around the unconditional mean with a width determined by the unconditional variance.

Q: Does this formula account for the uncertainty in estimating model parameters $ c $, $ \phi $, and $ \sigma_{\epsilon}^2 $?

No, this formula assumes that the model parameters $ c $, $ \phi $, and $ \sigma_{\epsilon}^2 $ are known true values. In practice, these parameters are estimated from data, introducing an additional source of uncertainty. For finite samples, accounting for parameter estimation uncertainty would typically result in wider prediction intervals, often requiring the use of t-distributions instead of z-distributions, or more complex bootstrap methods, especially for smaller sample sizes.

Question 1

Why is $ z_{\alpha/2} $ used, and why is it approximately 1.96 for a 95% interval?

Accepted Answer

The $ z_{\alpha/2} $ value comes from the standard normal distribution. When constructing a $ (1-\alpha) \times 100\% $ prediction interval, we want to capture the central $ (1-\alpha) $ portion of the distribution. This means $ \alpha/2 $ probability is left in each tail. For a 95% interval, $ \alpha = 0.05 $, so $ \alpha/2 = 0.025 $. The value $ z_{0.025} $ is the point such that $ P(Z > z_{0.025}) = 0.025 $ for a standard normal random variable $ Z $, which is approximately 1.96.

Question 2

How does the autoregressive parameter $ \phi $ influence the width of the prediction interval?

Accepted Answer

The parameter $ \phi $ determines the persistence of the series. If $ \phi $ is close to 0, the process is closer to white noise, and the interval width quickly stabilizes to $ \pm z_{\alpha/2} \sigma_{\epsilon} $. If $ |\phi| $ is large (closer to 1), the influence of past values is stronger, and the uncertainty accumulates more rapidly over time, leading to a much wider prediction interval for larger $ h $.

Question 3

What happens to the prediction interval as $ h $ approaches infinity for a stationary AR(1) process?

Accepted Answer

For a stationary AR(1) process (where $ |\phi| < 1 $), as $ h \to \infty $, the point forecast $ \hat{Y}_{t+h|t} $ converges to the unconditional mean of the process, $ E[Y_t] = c / (1-\phi) $. The sum $ \sum_{j=0}^{h-1} \phi^{2j} $ converges to $ 1 / (1-\phi^2) $. Thus, the forecast error variance converges to $ \sigma_{\epsilon}^2 / (1-\phi^2) $, which is the unconditional variance of the AR(1) process, $ Var(Y_t) $. The prediction interval will stabilize around the unconditional mean with a width determined by the unconditional variance.

Question 4

Does this formula account for the uncertainty in estimating model parameters $ c $, $ \phi $, and $ \sigma_{\epsilon}^2 $?

Accepted Answer

No, this formula assumes that the model parameters $ c $, $ \phi $, and $ \sigma_{\epsilon}^2 $ are known true values. In practice, these parameters are estimated from data, introducing an additional source of uncertainty. For finite samples, accounting for parameter estimation uncertainty would typically result in wider prediction intervals, often requiring the use of t-distributions instead of z-distributions, or more complex bootstrap methods, especially for smaller sample sizes.

Derivation of the Formula for a 95% Prediction Interval for an AR(1) Forecast

Visualizing...

The Formal Theorem

Analytical Intuition.

Institutional Warning.

Academic Inquiries.

Why is $z_{\alpha/2}$ used, and why is it approximately 1.96 for a 95% interval?

How does the autoregressive parameter $\phi$ influence the width of the prediction interval?

What happens to the prediction interval as $h$ approaches infinity for a stationary AR(1) process?

Does this formula account for the uncertainty in estimating model parameters $c$ , $\phi$ , and $\sigma_{\epsilon}^2$ ?

Standardized References.

Proof that Autocovariance Depends Only on Lag for Weakly Stationary Processes

Derivation of the Autocorrelation Function (ACF) for a White Noise Process

Proof of the Stationarity Condition for an AR(1) Process (|φ| < 1)

Proof of the Invertibility Condition for an MA(1) Process (|θ| < 1)

Institutional Citation

Dominate the Logic.