Define autocorrelation of order k and correlogram. For an infinite series generated by the moving average of random components with equal weights, obtain the correlogram of a moving average of extent m.
15 marks · directive: Obtain
The core of a good answer
The autocorrelation of order k is ρₖ = Cov(Xₜ, Xₜ₊ₖ)/Var(Xₜ) (sample rₖ = Σ(xₜ − x̄)(xₜ₊ₖ − x̄)/Σ(xₜ − x̄)²) and the correlogram is the graph of ρₖ (or rₖ) against k; for Xₜ = (1/m)(εₜ + εₜ₋₁ + ... + εₜ₋ₘ₊₁) with uncorrelated ε of variance σ², Cov(Xₜ, Xₜ₊ₖ) = (m − k)σ²/m², so ρₖ = 1 − k/m for 0 ≤ k ≤ m and ρₖ = 0 for k ≥ m: the correlogram is a straight line falling from 1 to 0 at lag m and zero thereafter.
An outline written by upsc.help, not an official UPSC answer; every fact checked by a separate review. How we work
What the question asks
Two definitions, then a derivation: count the overlapping random terms between Xₜ and Xₜ₊ₖ, divide by the variance, and describe the shape of the correlogram. Mention the implication for reading smoothed series.
Answer outline
Introduction
Let εₜ be a purely random series: E(εₜ) = 0, Var(εₜ) = σ², Cov(εₜ, εₛ) = 0 for t ≠ s. The moving average of extent m with equal weights is Xₜ = (1/m) Σⱼ₌₀ᵐ⁻¹ εₜ₋ⱼ.
Step 1: Definitions
- For a stationary series, γₖ = Cov(Xₜ, Xₜ₊ₖ) and ρₖ = γₖ/γ₀; ρ₀ = 1, ρ₋ₖ = ρₖ, |ρₖ| ≤ 1.
- Sample: rₖ = Σₜ₌₁ⁿ⁻ᵏ (xₜ − x̄)(xₜ₊ₖ − x̄) / Σₜ₌₁ⁿ (xₜ − x̄)².
- Correlogram: plot of rₖ against k = 1, 2, 3, ...; used to identify randomness (all rₖ near 0, within ±2/√n), trend (slow decay), oscillation (alternating signs) and MA or AR structure.
Step 2: Variance
- γ₀ = Var(Xₜ) = (1/m²) × m σ² = σ²/m.
Step 3: Covariance at lag k
- Xₜ contains εₜ, ..., εₜ₋ₘ₊₁; Xₜ₊ₖ contains εₜ₊ₖ, ..., εₜ₊ₖ₋ₘ₊₁.
- Common terms: εₜ₊ₖ₋ₘ₊₁, ..., εₜ, numbering m − k when 0 ≤ k < m, none when k ≥ m.
- Cross products of different ε have zero expectation, so γₖ = (m − k)σ²/m² for k < m and 0 for k ≥ m.
Step 4: Correlogram
- ρₖ = γₖ/γ₀ = (m − k)/m = 1 − k/m, k = 0, 1, ..., m; ρₖ = 0, k ≥ m.
- Example m = 5: ρ₁ = 0·8, ρ₂ = 0·6, ρ₃ = 0·4, ρ₄ = 0·2, ρ₅ = 0.
- Graph: points on the straight line from (0, 1) to (m, 0), then on the k-axis.
Step 5: Interpretation
- Averaging a random series creates positive correlation between neighbouring values even though the original terms are independent.
- The series is an MA(m − 1) process, so its autocorrelation cuts off after lag m − 1.
Conclusion
For an equal-weight moving average of extent m of random components, ρₖ = 1 − k/m up to lag m and zero beyond, a linearly declining correlogram that is the signature of a finite moving average.
Examples and data to cite
- In the Box-Jenkins method the autocorrelation cut-off after lag q is used to choose the order of an MA(q) model.
Facts to remember
- ρₖ = γₖ/γ₀; correlogram plots ρₖ or rₖ against k.
- For equal-weight MA of extent m: ρₖ = 1 − k/m (k ≤ m), 0 (k ≥ m).
- Approximate 95% band for white noise: ±2/√n.
- A moving average process of order q has ρₖ = 0 for k > q.