Persistence And Boundary Cases In AR(1) Models

Why the AR(1) Coefficient Matters

An AR(1) process uses the previous observation to describe the next one. Its coefficient, often written \( \phi \), indicates both how strongly deviations carry forward and whether they tend to keep or reverse their sign. A positive coefficient produces dependence in the same direction; a negative coefficient produces alternating dependence. The magnitude determines how quickly that effect fades, if it fades at all.

The key boundary is \( |\phi|=1 \). When \( |\phi|<1 \), deviations decay and the process can have a stable, time-invariant variance. At \( \phi=1 \), shocks do not decay; this is the unit-root case related to a random walk. At \( \phi=-1 \), shocks alternate without decaying. When \( |\phi|>1 \), shocks grow in magnitude. These distinctions matter when interpreting forecasts and deciding whether a stable time-series description is appropriate.

Definition: Consider the AR(1) equation \(X_t=c+\phi X_{t-1}+\varepsilon_t\), where \(c\) and \(\phi\) are constants and the innovations \(\varepsilon_t\) have mean zero and finite variance \(\sigma^2\). For second-order calculations, assume the innovations are uncorrelated over time and are uncorrelated with the process history. The process is weakly stationary when its mean and variance are finite and constant over time, and its covariance between two observations depends only on their time separation. Under this model, the stationary condition is \( |\phi|<1 \). If the innovations are independent and identically distributed, the stationary solution is also strictly stationary.

Sign, Magnitude, and Persistence

For a stationary AR(1), it is useful to write the equation in terms of its mean \( \mu \):

X_t-\mu=\phi(X_{t-1}-\mu)+\varepsilon_t, \qquad \mu=\frac{c}{1-\phi}.

This centered form makes persistence easier to interpret. If the process is currently \(d\) units above its mean, its expected deviation one period later is \( \phi d \); after \(k\) periods, the expected deviation is \( \phi^k d \). Thus, when \( \phi>0 \), the expected deviation keeps its sign and shrinks for \(0<\phi<1\). When \( \phi<0 \), its sign alternates while its magnitude shrinks if \(-1<\phi<0\). A coefficient close to zero means little persistence; one close to either boundary means slow decay.

The ACF of a stationary AR(1) is \( \rho_k=\phi^k \) at lag \(k\). The sign therefore alternates at successive lags for a negative coefficient, while the absolute correlation decays geometrically at rate \( |\phi| \). A useful summary is the half-life: the number of periods required for the magnitude of an expected deviation to fall to one-half of its starting value. For \(0<|\phi|<1\), solving \( |\phi|^h=0.5 \) gives:

h=\frac{\log(0.5)}{\log(|\phi|)}.

The half-life need not be an integer. It describes the continuous-valued number of periods at which the decay envelope reaches one-half; for negative \( \phi \), it does not remove the alternating sign. The formula does not apply at \( \phi=0 \), where the effect disappears after one step, or when \( |\phi|\geq 1 \), where there is no decaying envelope.

Worked Example: Comparing Persistent and Alternating Dependence

A hypothetical monthly index is modeled by a stationary AR(1) process with mean \( \mu=12 \), innovation distribution \( \varepsilon_t\sim N(0,1) \), and either \( \phi=0.8 \) or \( \phi=-0.8 \). For each model, calculate the stationary variance, the lag-3 autocorrelation, and the half-life. Explain how the two dependence patterns differ.

Method. For a stationary AR(1), the variance is \( \sigma^2/(1-\phi^2) \), the lag-\(k\) autocorrelation is \( \phi^k \), and the half-life is \( \log(0.5)/\log(|\phi|) \). Both coefficients satisfy \( |\phi|<1 \), so these stationary formulas apply.

Variance for either model. Here \( \sigma^2=1 \) and \( \phi^2=0.8^2=0.64 \). Therefore, \( \operatorname{Var}(X_t)=1/(1-0.64)=1/0.36=2.7778 \), rounded to four decimal places.

Lag-3 autocorrelation for \( \phi=0.8 \). Calculate \(0.8^3=0.8\times0.8\times0.8=0.512\). The correlation is positive, so deviations tend to retain their sign.

Lag-3 autocorrelation for \( \phi=-0.8 \). Calculate \((-0.8)^3=-0.512\). The magnitude is the same, but the correlation is negative at lag 3 because the sign alternates each period.

Half-life. For both models, the magnitude of the coefficient is \(0.8\), so \(h=\log(0.5)/\log(0.8)=3.1063\) periods, rounded to four decimals. The sign changes the direction of the deviations, not the rate at which their magnitude decays.

Check. Both models have the same variance and the same absolute lag-3 correlation. Their lag-3 correlations have opposite signs, as expected from the sign of \( \phi^3 \). The common mistake this comparison avoids is interpreting a negative coefficient as weak dependence merely because correlations alternate: its persistence is measured by \( |\phi| \).

Why the Stationarity Boundary Is at One

The variance condition can be seen directly by expanding the centered AR(1) equation. For clarity, let \(Z_t=X_t-\mu\), so \(Z_t=\phi Z_{t-1}+\varepsilon_t\). Repeated substitution gives a finite sum of past shocks and an initial-value term:

Z_t=\phi^t Z_0+\sum_{j=0}^{t-1}\phi^j\varepsilon_{t-j}.

To see why this determines the boundary, first take \(Z_0=0\) and assume the innovations are uncorrelated with variance \( \sigma^2>0 \). Because uncorrelated terms have zero covariance, the variance of the sum is the sum of the individual variances:

\operatorname{Var}(Z_t) =\sigma^2\sum_{j=0}^{t-1}\phi^{2j} = \begin{cases} \sigma^2\dfrac{1-\phi^{2t}}{1-\phi^2}, & \phi^2\neq 1,\\[6pt] t\sigma^2, & \phi^2=1. \end{cases}

When \( |\phi|<1 \), the powers \( \phi^{2t} \) tend to zero, and the variance approaches the finite value \( \sigma^2/(1-\phi^2) \). The dependence on the initial value also fades because \( \phi^t Z_0 \) tends to zero. This permits a stable stationary solution built from the indefinitely long history of shocks. When \( |\phi|=1 \), the variance in this finite-start calculation is \(t\sigma^2\), which grows without bound. When \( |\phi|>1 \), the powers grow, so the variance grows geometrically. Neither boundary case can have the finite, constant variance required for weak stationarity.

This argument assumes positive innovation variance. If \( \sigma^2=0 \), there are no random shocks and special deterministic sequences can behave differently; the usual AR(1) stationarity condition is stated for a nondegenerate innovation process. In the usual model with \( \sigma^2>0 \), equality at either boundary is not stationary.

Check the magnitude.
If \( |\phi|<1 \), shocks decay; if \( |\phi|=1 \), they persist without decay; if \( |\phi|>1 \), they grow.

Check the sign.
A positive coefficient preserves the sign of a deviation at each step. A negative coefficient reverses it.

Choose the interpretation.
Use stationary moments only when \( |\phi|<1 \). At or beyond the boundary, inspect the time-dependent variance and shock response instead.

Near the Boundary: Slow Decay Is Not a Unit Root

A coefficient close to one can look nearly nonstationary over a short observation period because its shocks fade slowly. However, a fixed coefficient such as \(0.95\) still satisfies the stationary condition: its effects eventually decay and its variance has a finite limit. This is distinct from \( \phi=1 \), where each shock remains in the level of the process indefinitely.

For \( \phi=1 \), the equation becomes \(X_t=c+X_{t-1}+\varepsilon_t\). If \(c=0\), this is a random walk; if \(c\neq0\), it is a random walk with drift. In either case, with positive innovation variance, the variance increases over time. For \( \phi=-1 \) and \(c=0\), the response to a shock alternates between positive and negative values of the same magnitude. The shock does not fade, and the variance also increases over time.

When \( |\phi|>1 \), the process is often called explosive: a one-unit shock has expected effect \( \phi^k \) after \(k\) periods, whose magnitude grows. “Explosive” describes the amplification of deviations; it does not guarantee that every observed path increases monotonically, since random innovations can change the direction of movement from one period to another.

Worked Example: A Stationary Process Close to the Unit-Root Boundary

A hypothetical quarterly measure follows a stationary AR(1) model with mean \( \mu=20 \), coefficient \( \phi=0.95 \), and innovations \( \varepsilon_t\sim N(0,1) \), independent over time. The latest observation is \(X_0=24\). Calculate the lag-10 autocorrelation, the half-life, and the expected value ten quarters ahead.

Method. Since \( |0.95|<1 \), the model is stationary. Use \( \rho_{10}=\phi^{10} \), the half-life formula, and the conditional forecast \(E[X_{10}\mid X_0]=\mu+\phi^{10}(X_0-\mu)\).

Lag-10 autocorrelation. Repeated multiplication gives \(0.95^{10}=0.5987369\), so \( \rho_{10}=0.5987 \) to four decimal places.

Half-life. Calculate \(h=\log(0.5)/\log(0.95)=13.5134\) quarters, rounded to four decimals. A value this close to one therefore implies slow decay, but not permanent persistence.

Ten-quarter forecast. The current deviation from the mean is \(24-20=4\). Its expected value after ten quarters is \(0.95^{10}\times4=0.5987369\times4=2.3949476\). Add the mean: \(E[X_{10}\mid X_0]=20+2.3949476=22.3949\), rounded to four decimal places.

Check. The forecast lies between the current value of 24 and the long-run mean of 20, as expected for a positive stationary coefficient below one. The common mistake is to call \(0.95\) a unit root because it is close to one. Its variance is finite: \(1/(1-0.95^2)=1/0.0975=10.2564\), rounded to four decimals.

Worked Example: Unit-Root, Alternating-Boundary, and Explosive Cases

Three hypothetical series start at \(X_0=0\), have zero intercept, and use independent innovations \( \varepsilon_t\sim N(0,1) \). Their respective coefficients are \(1\), \(-1\), and \(1.2\). Calculate each variance at time \(t=3\), and interpret the shock response.

Method. Expand \(X_3=\phi X_2+\varepsilon_3\) repeatedly. With \(X_0=0\), this gives \(X_3=\varepsilon_3+\phi\varepsilon_2+\phi^2\varepsilon_1\). Independence and unit innovation variance imply \( \operatorname{Var}(X_3)=1+\phi^2+\phi^4 \).

Coefficient \( \phi=1 \). The process is \(X_t=X_{t-1}+\varepsilon_t\). The time-3 value is \(X_3=\varepsilon_3+\varepsilon_2+\varepsilon_1\), so its variance is \(1+1+1=3\). Each shock remains in every later level.

Coefficient \( \phi=-1 \). The time-3 value is \(X_3=\varepsilon_3-\varepsilon_2+\varepsilon_1\). Its variance is \(1+(-1)^2+(-1)^4=1+1+1=3\). The shocks alternate in sign but do not shrink.

Coefficient \( \phi=1.2 \). The time-3 value is \(X_3=\varepsilon_3+1.2\varepsilon_2+1.44\varepsilon_1\). Its variance is \(1+1.2^2+1.44^2=1+1.44+2.0736=4.5136\). Earlier shocks receive increasingly large weights.

Check. All three cases have time-3 variance greater than the one-period innovation variance, but their mechanisms differ: nondecaying same-sign weights at \(1\), nondecaying alternating weights at \(-1\), and growing weights at \(1.2\). None is stationary. The common mistake this example avoids is treating \(-1\) as stationary because it is not greater than one; stationarity requires the strict inequality \( |\phi|<1 \).

Common Exam Traps and Practical Interpretation

Using \( \phi \) instead of \( |\phi| \) for decay. A negative coefficient can have slow persistence. The decay rate is determined by the magnitude; the sign determines whether dependence alternates.
Including the boundary in the stationary region. \( \phi=1 \) and \( \phi=-1 \) both fail the strict condition. With positive innovation variance, each produces time-varying variance.
Calling a near-unit coefficient a unit root. A coefficient such as \(0.95\) is stationary, although it can be highly persistent. Exactly \(1\) is the unit-root boundary.
Using stationary variance outside its assumptions. The formula \( \sigma^2/(1-\phi^2) \) requires \( |\phi|<1 \). Substituting \( |\phi|\geq1 \) does not provide a valid stationary variance.
Reading explosive behavior as a guaranteed upward path. A coefficient above one amplifies deviations and shocks in magnitude; random innovations can still make individual observations rise or fall.

In an applied time-series analysis, distinguishing slow decay from a unit root is important because the two imply different long-horizon behavior. A stationary model eventually forgets a shock, even if that takes many periods. A unit-root model retains it, and an explosive model amplifies it. A short series may not make the distinction visually obvious, so interpretation must follow the model’s coefficient and assumptions rather than appearance alone.

Key takeaway: For an AR(1), \( |\phi|<1 \) means stationary decay; the sign of \( \phi \) determines whether dependence keeps or alternates its sign. The cases \( \phi=1 \) and \( \phi=-1 \) are nonstationary boundaries, while \( |\phi|>1 \) is explosive.

Check Your Understanding

For each problem, identify the relevant persistence or boundary behavior and show the calculation before checking the final answer.

A stationary AR(1) process has \( \phi=0.5 \) and innovation variance \( \sigma^2=1 \). Find its stationary variance and lag-3 autocorrelation. (Answer: variance \(4/3\); lag-3 autocorrelation \(0.125\).)
A stationary AR(1) process has \( \phi=-0.8 \) and innovation variance \( \sigma^2=0.36 \). Find its stationary variance and lag-4 autocorrelation, and state whether the lag-4 correlation is positive or negative. (Answer: variance \(1\); lag-4 autocorrelation \(0.4096\), positive.)
A stationary AR(1) process has \( \phi=0.9 \). Calculate its half-life to four decimal places and give the first whole number of periods by which the decay magnitude is below one-half. (Answer: half-life \(6.5788\) periods; below one-half after 7 periods.)
A process starts at \(X_0=0\) and follows \(X_t=-X_{t-1}+\varepsilon_t\), with independent \( \varepsilon_t\sim N(0,1) \). Find \( \operatorname{Var}(X_4) \) and classify the process. (Answer: variance \(4\); nonstationary boundary case with alternating, nondecaying shocks.)
A process starts at \(X_0=0\) and follows \(X_t=1.2X_{t-1}+\varepsilon_t\), with independent \( \varepsilon_t\sim N(0,1) \). Find \( \operatorname{Var}(X_3) \), and state whether the process is stationary, unit-root, or explosive. (Answer: variance \(4.5136\); explosive and nonstationary.)

What You'll Learn