STAT8003 Chap.4 Linear Processes and Moving Averages
Linear Processes and Moving Averages
Build the innovation representation
This chapter defines linearity, causality and the general innovation representation. A linear process uses weighted shocks; a causal one excludes future shocks; a stationary causal linear process becomes the general linear process used for moment calculations.
Square-summable weights keep the process variance finite even when the history is infinite.
Turn weights into moments
Variance adds squared coefficients, while lagged covariance aligns two copies of the weight sequence and adds products at shared innovation positions. This single method covers the general formulas and the finite MA cases.
It is safer than recalling a numerator because it retains the sign attached to each coefficient.
Read finite memory from the ACF
An MA(q) value contains only the current innovation and q past innovations. Values farther than q periods apart share none, so the population ACF cuts off after lag q. Sample correlations fluctuate around this structure rather than printing exact zeros.
The course writes MA coefficients with minus signs, so the visible equation must be translated before a remembered formula is used.
What this chapter covers
- 01
Linear process
- 02
Causal representation
- 03
Psi weights
- 04
MA model convention
- 05
MA moments
- 06
ACF cutoff
Worked example · free
Derive an MA(1) correlation
- 3Square and add the innovation coefficients.
- 2Multiply coefficients on the shared lag-one shock.
- 2Standardise the covariance.
Key terms
- Linear Process
- A process represented as a linear function of an innovation sequence. Its mean, variance and lagged covariance can be derived directly from the innovation weights and their aligned products.
- Causal Process
- A process depending only on innovations available at the current or earlier times. This direction makes the representation usable for real-time interpretation and forecasting without future shocks.
- ACF Cutoff
- The population autocorrelation becoming exactly zero after a finite lag. For an MA(q), the cutoff occurs after q because more widely separated values share no innovations.
Linear Processes and Moving Averages FAQ
Why is every finite MA process stationary?
An MA model is a finite linear combination of identically distributed innovations with fixed coefficients. Its mean and variance are constant, and its covariance is determined by which coefficients overlap at each lag.
Why does the MA autocorrelation cut off?
Values more than q periods apart in an MA(q) share no innovations. Independence then makes their covariance zero, so the population ACF vanishes after lag q even though sample correlations will fluctuate around zero.
What does the back-shift operator add?
It compresses lag notation. Powers of B shift a series backward by the matching number of periods, allowing an MA model to be written as a characteristic polynomial applied to the innovation sequence.
How do I avoid an MA sign-convention mistake?
Read coefficients from the written equation before using a formula. In this course, the MA terms are written with minus theta coefficients, so a visible plus sign corresponds to a negative theta. Rebuild covariance by aligning signed innovation coefficients and use the one-half magnitude bound for an MA(1) correlation as a final check. A borrowed formula from a plus-sign convention can otherwise reverse the result.
What does square summability protect?
It keeps the variance of an infinite linear innovation representation finite by requiring the squared weights to have a finite sum. Finite-order MA models satisfy this automatically because all weights after q are exactly zero.
Exam move
Rebuild MA formulas from coefficient overlap instead of memorising them in isolation. Write the innovation coefficients in rows, shift one row by the required lag and multiply only aligned entries. Then divide by the variance for autocorrelation. Finish by checking the expected cutoff and confirming that the sign follows this course's minus-theta convention.
Start with a general linear process and explain what each adjective adds: linearity supplies a weighted innovation representation, causality removes future shocks, and stationarity stabilises the moments. Check square summability before claiming that an infinite representation has finite variance. For MA work, translate every written equation into a coefficient list headed lag zero through lag q.
Preserve the actual signs from the equation. At lag k, shift a second copy by k positions, multiply the aligned coefficients and add. This procedure derives gamma zero, gamma one and every later covariance without relying on a convention that may differ across books. Draw the population ACF support and identify the last possible nonzero lag. Then explain why a sample ACF will show noisy bars rather than exact zeros.
Practise the back-shift operator by converting between an expanded model and theta(B)at in both directions, checking the exponent and sign of each term. Use two sanity checks: an MA(1) lag-one correlation cannot have magnitude above one half, and an MA(q) covariance must vanish beyond q. If either check fails, revisit coefficient alignment.
Finish by contrasting order with invertibility so that the finite innovation window is not mistaken for the reverse-representation condition introduced later.
Working through Linear Processes and Moving Averages in STAT8003? Sia is AskSia’s AI Statistics tutor — ask any STAT8003 Linear Processes and Moving Averages question and get a clear, step-by-step explanation grounded in how STAT8003 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.