Common Errors In AR Model Interpretation

Why AR Interpretations Need Care

An autoregressive model can be written in more than one algebraically equivalent form. Its intercept is not generally its mean, the signs in its characteristic polynomial depend on the model’s stated form, and a short run of positive autocorrelations does not establish that the full ACF stays positive. These details matter when deciding whether a model is stationary and when interpreting its dependence pattern.

This tutorial focuses on common interpretation errors for AR(1) and AR(2) models. Assume the innovations are white noise: they have mean zero, constant variance \(\sigma_\varepsilon^2\), and zero autocovariance at nonzero lags. For the usual stationary AR results below, also assume each innovation is uncorrelated with past values of the process.

Definition: In the convention used here, an AR(1) model is \(Y_t=c+\phi Y_{t-1}+\varepsilon_t\). An AR(2) model is \(Y_t=c+\phi_1Y_{t-1}+\phi_2Y_{t-2}+\varepsilon_t\). The parameters \(c\), \(\phi\), \(\phi_1\), and \(\phi_2\) are coefficients in these equations; \(c\) is called the intercept.

Intercept Is Not Usually the Mean

For a stationary AR process with a finite mean, the mean is constant over time. Taking expectations in the AR(1) equation gives \(\mu=c+\phi\mu\), because the innovation has mean zero. Thus, when \(\phi\ne1\), the mean is \(c/(1-\phi)\). The same calculation for an AR(2) gives \(\mu=c/(1-\phi_1-\phi_2)\), provided the model is stationary.

These formulas do not make a nonstationary model stationary. In particular, a finite-looking mean formula is not a substitute for checking the stationarity conditions.

\text{AR(1): }\mu=\frac{c}{1-\phi} \qquad \text{AR(2): }\mu=\frac{c}{1-\phi_1-\phi_2}

Worked Example: Finding an AR(1) Mean and ACF

A process is specified as \(Y_t=0.6+0.7Y_{t-1}+\varepsilon_t\), where the innovations are white noise with variance \(4\). Determine whether the process is stationary, then find its mean, variance, and lag-3 autocorrelation.

Method. First check the AR(1) stationarity condition. For an AR(1), stationarity requires \(|\phi|<1\). Here \(|0.7|<1\), so the stationary AR(1) formulas apply.

Mean. The intercept is \(c=0.6\), not the mean. Therefore \[ \mu=\frac{c}{1-\phi}=\frac{0.6}{1-0.7}=\frac{0.6}{0.3}=2. \] A check using the mean equation gives \(0.6+0.7(2)=2\), as required.

Variance. For a stationary AR(1), \(\operatorname{Var}(Y_t)=\sigma_\varepsilon^2/(1-\phi^2)\). Hence \[ \operatorname{Var}(Y_t)=\frac{4}{1-0.7^2} =\frac{4}{1-0.49} =\frac{4}{0.51} \approx 7.8431. \]

Lag-3 autocorrelation. The AR(1) ACF is \(\rho_k=\phi^k\), so \(\rho_3=0.7^3=0.343\). The final answers are mean \(2\), variance approximately \(7.8431\), and lag-3 autocorrelation \(0.343\). The variance is rounded to four decimal places. This calculation avoids the common mistake of using the intercept \(0.6\) as the process mean.

Stationarity and Coefficient Signs

For an AR(1), the familiar condition \(|\phi|<1\) is both necessary and sufficient for weak stationarity under the white-noise assumptions. For an AR(2) in the convention stated above, stationarity requires all three inequalities below. They are not equivalent to checking only whether the individual coefficients have absolute values below one.

Theorem: The AR(2) process \(Y_t=c+\phi_1Y_{t-1}+\phi_2Y_{t-2}+\varepsilon_t\) is stationary if and only if \[ -1<\phi_2<1,\qquad \phi_1+\phi_2<1,\qquad \phi_2-\phi_1<1. \] Equivalently, the roots of \(1-\phi_1z-\phi_2z^2=0\) must lie outside the unit circle.

The polynomial’s signs follow directly from the model convention: move the lag terms to the left to obtain \(Y_t-\phi_1Y_{t-1}-\phi_2Y_{t-2}=\) intercept plus innovation. Its AR polynomial is therefore \(1-\phi_1z-\phi_2z^2\). If the model is written using minus signs in front of its coefficients, translate it into the stated convention before applying a stationarity test.

Worked Example: Checking AR(2) Stationarity and Dependence

Consider \(Y_t=0.15+1.2Y_{t-1}-0.5Y_{t-2}+\varepsilon_t\), with white-noise innovations. Check stationarity and find the mean and the first four autocorrelations.

Stationarity check. Here \(\phi_1=1.2\) and \(\phi_2=-0.5\). The three tests are: \[ -1<-0.5<1,\qquad 1.2+(-0.5)=0.7<1,\qquad -0.5-1.2=-1.7<1. \] All conditions hold, so the process is stationary.

Mean. The mean is \[ \mu=\frac{0.15}{1-1.2-(-0.5)} =\frac{0.15}{0.3}=0.5. \] Substitution checks the result: \(0.15+1.2(0.5)-0.5(0.5)=0.5\).

Autocorrelations. For a stationary AR(2), the Yule–Walker equations give \(\rho_1=\phi_1/(1-\phi_2)\), and for \(k\ge2\), \(\rho_k=\phi_1\rho_{k-1}+\phi_2\rho_{k-2}\), with \(\rho_0=1\). Thus \[ \rho_1=\frac{1.2}{1-(-0.5)}=\frac{1.2}{1.5}=0.8, \] \[ \rho_2=1.2(0.8)-0.5(1)=0.46, \quad \rho_3=1.2(0.46)-0.5(0.8)=0.152, \] \[ \rho_4=1.2(0.152)-0.5(0.46)=-0.0476. \] The sequence starts positive but is not positive at every lag. Indeed, the recurrence has characteristic equation \(\lambda^2-1.2\lambda+0.5=0\), whose roots are complex, with modulus \(\sqrt{0.5}<1\); this is consistent with damped oscillation.

For a stationary AR(2), the theoretical PACF is zero at lags greater than 2, whereas the ACF generally continues rather than cutting off at lag 2. The answer is a stationary process with mean \(0.5\) and autocorrelations \(0.8,\ 0.46,\ 0.152,\ -0.0476\) at lags 1 through 4. The calculation avoids inferring the entire ACF’s sign from its first few values.

How the ACF and PACF Results Fit Together

The AR(1) ACF result can be justified from the model. Subtract the stationary mean and write \(Z_t=Y_t-\mu\), so \(Z_t=\phi Z_{t-1}+\varepsilon_t\). For \(k\ge1\), take covariance with \(Z_{t-k}\). The innovation is uncorrelated with the past, so \(\operatorname{Cov}(\varepsilon_t,Z_{t-k})=0\), giving \(\gamma_k=\phi\gamma_{k-1}\). Dividing by \(\gamma_0\) yields \(\rho_k=\phi\rho_{k-1}\). Since \(\rho_0=1\), repeated substitution gives \(\rho_k=\phi^k\).

This result explains why a positive AR(1) coefficient gives a positive ACF that decays geometrically, while a negative coefficient gives alternating signs. It does not mean the PACF has the same pattern: for an AR(1), the theoretical PACF cuts off after lag 1. For a stationary AR(2), the corresponding cutoff is after lag 2. A cutoff is a statement about the theoretical PACF, not a claim that estimated sample correlations are exactly zero.

Worked Example: Rejecting an Incomplete AR(2) Stationarity Check

A proposed model is \(Y_t=0.2+0.8Y_{t-1}+0.4Y_{t-2}+\varepsilon_t\). A reviewer says it is stationary because \(|\phi_2|=0.4<1\). Check the claim and identify the coefficient-sign trap in forming the polynomial.

Method. Apply all AR(2) conditions, not just the bound on \(\phi_2\). Here \(\phi_1=0.8\) and \(\phi_2=0.4\). Although \(-1<0.4<1\), the second condition fails: \[ \phi_1+\phi_2=0.8+0.4=1.2\not<1. \] The third condition holds because \(\phi_2-\phi_1=0.4-0.8=-0.4<1\). Since one required condition fails, the process is not stationary.

Polynomial check. The model convention gives the polynomial \(1-0.8z-0.4z^2\), not \(1+0.8z+0.4z^2\). Solving \(1-0.8z-0.4z^2=0\) gives \[ z=\frac{-0.8\pm\sqrt{0.8^2+4(0.4)(1)}}{2(0.4)} =\frac{-0.8\pm\sqrt{2.24}}{0.8}. \] The roots are approximately \(0.8708\) and \(-2.8708\). Because one root lies inside the unit circle, the root condition also fails. The conclusion is nonstationarity. The reviewer’s conclusion uses an incomplete test, and using the wrong polynomial signs would be a separate error.

Common Interpretation Errors

Calling the intercept the mean. In an AR(1), use \(c/(1-\phi)\); in an AR(2), use \(c/(1-\phi_1-\phi_2)\), and only interpret these as stationary means when stationarity holds.
Checking only one AR(2) inequality. All three conditions must hold. A coefficient such as \(\phi_2\) being between \(-1\) and \(1\) is not enough.
Changing coefficient signs without changing conventions. For the plus-coefficient form, the AR(2) polynomial is \(1-\phi_1z-\phi_2z^2\). Write the equation and derive the polynomial rather than guessing its signs.
Confusing ACF and PACF behavior. A stationary AR(\(p\)) has a PACF cutoff after lag \(p\); its ACF generally tails off. Do not reverse these patterns.
Reading a theoretical pattern literally from a sample plot. Sampling variability means estimated ACF and PACF values need not exactly match their theoretical counterparts.

These checks are useful when interpreting a fitted model, comparing a proposed coefficient set with stationarity requirements, or reading an ACF/PACF display. A careful interpretation starts with the model’s exact equation, verifies stationarity, and only then applies stationary mean and dependence results. The distinctions are aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.

Check Your Understanding

For each problem, apply the model convention and conditions stated in the question before interpreting the result.

A stationary AR(1) is \(Y_t=0.9+0.4Y_{t-1}+\varepsilon_t\), with innovation variance \(2\). Find its mean and lag-2 autocorrelation. (Answer: mean \(1.5\); lag-2 ACF \(0.16\))
For \(Y_t=0.1+0.5Y_{t-1}+0.6Y_{t-2}+\varepsilon_t\), check all AR(2) stationarity inequalities. (Answer: nonstationary; \(\phi_1+\phi_2=1.1\) fails the required inequality)
A stationary AR(2) has \(\phi_1=0.6\), \(\phi_2=-0.2\), and intercept \(c=0.3\). Find its mean and \(\rho_1\). (Answer: mean \(0.5\); \(\rho_1=0.5\))
A stationary AR(2) has \(\rho_0=1\), \(\rho_1=0.4\), \(\phi_1=0.7\), and \(\phi_2=-0.3\). Use the AR(2) recurrence to calculate \(\rho_2\) and \(\rho_3\). (Answer: \(\rho_2=-0.02\); \(\rho_3=-0.134\))
A colleague claims that a stationary AR(2) must have ACF equal to zero after lag 2 because its PACF cuts off there. State which function cuts off and whether the claim is correct. (Answer: the PACF cuts off after lag 2; the claim about the ACF is incorrect because it generally tails off)

What You'll Learn