ECON31031 Chap.7 AR, MA and ARMA Processes
AR, MA and ARMA Processes
AR, MA and ARMA Processes are the stationary models of Part II. An AR(1) is stationary when its coefficient lies strictly between minus one and one; then its mean is the constant divided by one minus the coefficient, its variance is the shock variance divided by one minus the squared coefficient, and its autocorrelations decay geometrically.
A moving average of order q has autocorrelations that cut off after lag q and should be invertible. The partial autocorrelation function shows the mirror image, cutting off after p lags for an AR(p). Together these give the eyeball rules for identification. ARMA models combine both parts parsimoniously, but stationarity conditions become joint for higher orders, as the AR(2) inequalities and the characteristic equation show.
The supplementary reading derives the AR(1) moments by repeated substitution, writing the process as an infinite moving average whose weights are powers of the coefficient. The same idea underlies Wold's theorem and explains why ARMA models can approximate a wide range of stationary behaviour with only a few parameters.
What this chapter covers
- 01
AR(1) definition and stationarity condition
- 02
Mean, variance and ACF of an AR(1)
- 03
AR(p) processes and lag polynomials
- 04
MA(q) moments, cut-offs and invertibility
- 05
The PACF and the eyeball identification rules
- 06
Wold's theorem and ARMA(p,q)
- 07
AR(2) stationarity inequalities and characteristic roots
Worked example · free
Derive the moments of an AR(1) and an MA(1)
- 2AR(1) mean: 3/(1 - 0.5) = 6. Variance: 3/(1 - 0.25) = 4.
- 2Second autocorrelation: 0.5 squared = 0.25, and the pattern keeps halving.
- 2MA(1): first autocorrelation = -0.4/(1 + 0.16) = -0.345, and zero beyond lag 1.
Key terms
- Invertibility
- The property of a moving average whose lag polynomial has an inverse, so it can be rewritten as an autoregression.
- Characteristic equation
- The polynomial equation whose roots decide whether an autoregression is stationary.
- Partial autocorrelation
- The last coefficient of an autoregression of a given order, viewed as a function of that order.
- Wold's theorem
- The result that every stationary process has an infinite moving average form.
AR, MA and ARMA Processes FAQ
How do I tell an AR from an MA in a correlogram?
Use the eyeball rules from the notes. An AR(p) has an ACF that decays and a PACF that cuts off after p lags. An MA(q) has an ACF that cuts off after q lags and a PACF that decays. Mixed patterns suggest an ARMA.
When is an AR(2) stationary?
All three inequalities must hold: the two coefficients sum to less than one, the second minus the first is less than one, and the second is greater than minus one. Equivalently, the roots of the characteristic equation must all be smaller than one in absolute value.
Why use an ARMA instead of a long AR?
An ARMA can reproduce a rich autocorrelation pattern with fewer parameters than a long pure AR or MA. That parsimony is why classical forecasting methods are built around ARMA models, even though estimation becomes harder.
What does invertibility mean for an MA process?
It means the moving average lag polynomial has an inverse, so the process can be rewritten as an autoregression in the observed series plus a white noise shock. Invertibility restricts the MA coefficients in the same way that stationarity restricts the AR coefficients.
Why does the PACF cut off for an AR(p)?
The PACF at lag k is the last coefficient from an autoregression of order k. If the true process is an AR(p), autoregressions of higher order add coefficients that are zero in the population, so the sample PACF should drop to near zero after lag p.
What does the sign of the AR(1) coefficient change?
A positive coefficient gives autocorrelations that stay positive and decline geometrically, so the series moves smoothly. A negative coefficient makes autocorrelations alternate in sign while shrinking at the same rate, so the series tends to oscillate from one period to the next.
Exam move
Derive the AR(1) moments from the infinite moving average form once, then memorise them. Practise MA(1) and MA(2) autocorrelations from the autocovariance definitions. Sketch the four eyeball patterns from memory: ACF and PACF for an AR and an MA. For AR(2) questions, always check both the inequalities and the characteristic roots, and state the verdict in words.
Keep a one-page identification sheet: the four ACF and PACF patterns, the AR(1) and MA(q) moment formulas, the AR(2) inequalities and the characteristic equation. Test yourself by covering the model and naming it from a described correlogram, then reverse the exercise and describe the correlogram from the model.
Working through AR, MA and ARMA Processes in ECON31031? Sia is AskSia’s AI Economics tutor — ask any ECON31031 AR, MA and ARMA Processes 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.