ECON31031 Chap.8 ARMA Forecasting, Estimation and Model Selection
ARMA Forecasting, Estimation and Model Selection
ARMA Forecasting, Estimation and Model Selection turns fitted models into forecasts and chooses between them. The vocabulary matters: in-sample and out-of-sample data, extrapolation, the forecast base and the horizon.
A one-step AR forecast is the conditional expectation given the last value; further steps feed forecasts back into the recursion when extrapolating and use actual values when they are available, and every stationary forecast converges to the unconditional mean. MA forecasts set unknown future shocks to zero.
Full ARMA models are estimated by maximum likelihood under normality, and their orders are chosen by minimising AIC or BIC, which penalise model dimension differently. Forecast evaluation uses the same vocabulary. Holding back an out-of-sample period lets you compare forecasts with outcomes, and when actual values are available they replace forecasts in the recursion.
Model selection then balances fit against size: the notes favour information criteria over sequences of significance tests because every ARMA is only an approximation.
What this chapter covers
- 01
Forecast vocabulary: in-sample, out-of-sample, base and horizon
- 02
Conditional and unconditional expectations
- 03
Multi-step AR(1) and AR(p) forecasts
- 04
MA(q) forecasts and mean reversion
- 05
Maximum likelihood estimation of ARMA models
- 06
AIC and BIC model selection
Worked example · free
Forecast an AR(1) three steps ahead
- 1One step: 1 + 0.8(10) = 9.0.
- 2Two steps: 1 + 0.8(9.0) = 8.2; three steps: 1 + 0.8(8.2) = 7.56.
- 2The long-run mean is 1/(1 - 0.8) = 5, so the forecasts fall towards 5 as the horizon grows.
Key terms
- Forecast horizon
- The number of periods ahead of the forecast base for which a forecast is made.
- Information criterion
- A measure of fit penalised by model dimension, minimised to choose the order of a model.
- Mean reversion
- The convergence of forecasts from a stationary model towards its unconditional mean as the horizon lengthens.
ARMA Forecasting, Estimation and Model Selection FAQ
Why do AR forecasts converge to the mean?
Each step multiplies the gap between the last forecast and the mean by the AR coefficient, which is smaller than one in absolute value for a stationary process. The gap therefore shrinks geometrically, so long-horizon forecasts approach the unconditional mean.
Why can AIC and BIC choose different models?
AIC adds two for each parameter, while BIC adds the log of the sample size. Once the sample exceeds about seven observations, the BIC penalty is larger, so BIC tends to choose smaller models than AIC.
Why is ARMA estimated by maximum likelihood?
The moving average part depends on unobserved past shocks, and least squares criteria are inefficient. Maximum likelihood under normality handles the shocks numerically, along with starting values and the stationarity and invertibility restrictions.
What is the forecast base?
It is the last date of the in-sample period, the point from which forecasts are made. A one-step forecast targets the next date after the base, and longer horizons count further periods ahead, such as four steps for an annual horizon with quarterly data.
How does an MA(1) forecast behave beyond one step?
The one-step forecast uses the last estimated shock multiplied by the MA coefficient. From two steps ahead, all relevant shocks are in the future and are set to zero, so the forecast equals the mean of the process for every longer horizon.
What is the difference between in-sample and out-of-sample data?
In-sample data are the observations used to estimate the model's parameters. Out-of-sample data are held back from estimation and used to evaluate forecasts, so that forecast accuracy is judged on observations the model has not already been fitted to.
Can I test ARMA terms with t statistics instead?
Testing individual terms is possible, but the notes prefer information criteria because every fitted ARMA is only an approximation to an infinite moving average. Choosing the model that minimises a penalised criterion balances fit against size more coherently than a chain of separate tests.
Exam move
Practise the forecast recursion on paper for AR(1), AR(2) and MA(1) models, writing beside each lag whether it is data or a forecast. Compute AIC and BIC for a small table of models, circle the minimum and explain any disagreement. Learn the forecasting vocabulary precisely, because short definition questions reward exact wording.
For selection questions, build a small table with columns for the log-likelihood, the parameter count, AIC and BIC, and fill it row by row. Practise explaining a disagreement between the two criteria in one sentence, because that explanation is often what distinguishes a complete answer.
Finally, practise one forecast question in which the out-of-sample value is supplied, so that switching between actual and forecast lags becomes automatic.
Working through ARMA Forecasting, Estimation and Model Selection in ECON31031? Sia is AskSia’s AI Economics tutor — ask any ECON31031 ARMA Forecasting, Estimation and Model Selection question and get a clear, step-by-step explanation grounded in how ECON31031 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.