Autocovariance and Lag Polynomials: Econometrics Tutorial 2025

Introduction to Autocovariance and Lag Polynomials

Econometrics students often encounter time series concepts like autocovariance, autocorrelation, and lag polynomials. These tools are essential for modeling dynamic relationships in data, whether you're analyzing stock returns, GDP growth, or even social media trends. In this tutorial, we'll break down the core ideas using a homework-style problem, connecting them to real-world examples like AI-driven forecasting or viral app adoption.

What Are Autocovariances and Why Do They Matter?

Autocovariance measures how a time series relates to its own past values. For a stationary process, it depends only on the lag between observations. This concept is crucial for understanding persistence in economic data, such as inflation rates or consumer spending. In our first example, we consider a moving average process: x_t = α0 + 5*u_t + u_{t-1}, where u_t is white noise with variance σ². To find the autocovariances, we compute γ(k) = Cov(x_t, x_{t+k}). For k=0, γ(0) = Var(x_t) = 25σ² + σ² = 26σ². For k=1, γ(1) = Cov(5u_t + u_{t-1}, 5u_{t+1} + u_t) = 5σ². For k>1, γ(k)=0. This shows that the process has memory only up to one period, similar to how a one-day trend in app downloads might fade quickly.

Autocovariances for an AR(1) Process

Now consider a stationary AR(1) process: x_t = 3 + 0.5*x_{t-1} + u_t, with Var(u_t)=3. The variance of x_t is σ_x² = σ_u²/(1-φ²) = 3/(1-0.25)=4. The first three autocovariances are γ(0)=4, γ(1)=φ*γ(0)=0.5*4=2, γ(2)=φ²*γ(0)=0.25*4=1, γ(3)=φ³*γ(0)=0.125*4=0.5. Autocorrelations are ρ(1)=0.5, ρ(2)=0.25, ρ(3)=0.125. This exponential decay is typical for AR(1) processes, similar to how the impact of a policy change gradually diminishes.

Variance Matrix for an AR(1) Process

Given e_t = a e_{t-1} + u_t with a=0.5, we compute the variance matrix for the vector E = (e1, e2, e3). Assuming e0 is fixed, the variance of e1 is Var(e1) = σ_u² = 3. Then Var(e2) = a²Var(e1) + σ_u² = 0.25*3+3=3.75, and Var(e3) = a²Var(e2)+σ_u² = 0.25*3.75+3=3.9375. Covariances: Cov(e1,e2)=aVar(e1)=0.5*3=1.5, Cov(e1,e3)=a²Var(e1)=0.25*3=0.75, Cov(e2,e3)=aVar(e2)=0.5*3.75=1.875. The variance matrix is symmetric with these values. If the process is stationary, Var(e_t)=σ_u²/(1-a²)=3/(0.75)=4 for all t, and Cov(e_t, e_{t+k})=a^k*4. Thus, the stationary variance matrix has all diagonal entries 4, off-diagonals decaying.

Lag Polynomials: Operations and Roots

Lag polynomials are powerful tools for representing time series models. Define a(L)=1-2L and b(L)=3-0.3L+0.5L². To compute a(L)x_t for given values: a(L)x_t = x_t - 2x_{t-1}. With x_t=3, x_{t-1}=-3, we get 3 - 2*(-3)=9. For b(L)x_t = 3x_t -0.3x_{t-1}+0.5x_{t-2}=3*3 -0.3*(-3)+0.5*(-2)=9+0.9-1=8.9. The roots of a(L) solve 1-2z=0 => z=0.5. For b(L), solve 3-0.3z+0.5z²=0 using quadratic formula: z = [0.3 ± sqrt(0.09 - 6)]/(1) = [0.3 ± sqrt(-5.91)]/1, complex roots. The product c(L)=a(L)b(L)= (1-2L)(3-0.3L+0.5L²) = 3 -0.3L+0.5L² -6L +0.6L² -L³ = 3 -6.3L +1.1L² -L³. Its roots are the union of roots of a and b: 0.5 and the complex pair. To invert b(L), we find b^{-1}(L) such that b(L)b^{-1}(L)=1. This involves solving for coefficients; typically, the inverse exists if roots are outside the unit circle. Here, since b has complex roots with modulus sqrt(0.09+6)/? Actually, modulus of roots = sqrt(0.09+6)/? Wait, compute modulus: for quadratic 0.5z² -0.3z+3=0, roots are complex with modulus sqrt(3/0.5)=sqrt(6)≈2.45, so invertible. The coefficients of b^{-1}(L) up to L² can be found by solving (3-0.3L+0.5L²)(c0 + c1L + c2L² + ...)=1. Equating coefficients: 3c0=1 => c0=1/3; 3c1 -0.3c0=0 => c1=0.1/3=0.03333; 3c2 -0.3c1+0.5c0=0 => c2 = (0.3c1 -0.5c0)/3 = (0.01 -0.1667)/3 = -0.05223. So b^{-1}(L) ≈ 0.3333 + 0.03333L -0.05223L² + ....

Practical Application: Autocorrelation Detection

In homework, you often regress residuals on lagged residuals to test for autocorrelation. For example, after estimating a consumption growth model, you can compute residuals and regress e_t on e_{t-1}. A significant coefficient indicates autocorrelation, which can bias standard errors. This is similar to checking if a stock's returns depend on yesterday's returns—a common phenomenon in financial econometrics.

Conclusion

Mastering autocovariance, variance matrices, and lag polynomials is essential for time series analysis. These concepts appear in advanced topics like VAR models, cointegration, and machine learning for time series. By practicing with concrete examples, you'll build intuition for econometric modeling. For more econometrics homework help and study guides, explore our other resources.