AR Forecasting

Forecasting workflow for autoregressive models: estimate parameters from data, then compute the best linear predictor and prediction intervals.

1. Parameter Estimation from Data

Given observations $\{x_t\}_{t=1}^n$ from an AR(1) model $X_t = \phi_1 X_{t-1} + Z_t$:

  1. Compute sample ACVF: $\hat{\gamma}_X(0)$ and $\hat{\gamma}_X(1)$
  2. Estimate $\hat{\phi}_1 = \hat{\gamma}_X(1) / \hat{\gamma}_X(0)$
  3. Estimate $\hat{\sigma}^2 = \hat{\gamma}_X(0)(1 - \hat{\phi}_1^2)$

Alternatively, fit via lm() regression of $(x_2, \dots, x_n)$ on $(x_1, \dots, x_{n-1})$ without intercept.

2. One-Step Forecast

$$P_n X_{n+1} = \hat{\phi}_1 x_n$$$$\text{MSE} = \hat{\sigma}^2$$

3. Prediction Interval

95% prediction interval:

$$P_n X_{n+1} \pm 1.96 \hat{\sigma}$$

4. Comparing Methods

When comparing the best linear predictor $P_n X_{n+1}$ (hand-computed from ACVF) with a software-based forecast (e.g., itsmr::forecast()), the results should be close but may differ slightly due to:

  • Different estimation methods
  • Classical decomposition accounting for trend/seasonality
  • Different assumed model structure