The Mean of an AR(1) Process

Why the AR(1) Mean Needs Careful Interpretation

An AR(1) model uses the previous observation to describe the current one. Its intercept, its average level over time, and its expected next value given the latest observation are related, but they are not generally the same quantity. Keeping those interpretations separate is important when describing a fitted model or making a forecast.

This tutorial focuses on the mean of an AR(1) process. It assumes familiarity with the AR(1) specification and with weak stationarity, introduced in the earlier material on those subjects. The distinction is aligned with the SOA Exam SRM (Statistics for Risk Modeling) syllabus.

Model and assumptions: Consider the AR(1) process $$ X_t=c+\phi X_{t-1}+\varepsilon_t, $$ where $c$ is a constant intercept, $\phi$ is the autoregressive coefficient, and $\varepsilon_t$ is an innovation with finite mean zero. For the stationary results below, assume the usual causal, weakly stationary AR(1) model with $|\phi|<1$. For the conditional-mean formula, also assume $E[\varepsilon_t\mid\mathcal{F}_{t-1}]=0$, where $\mathcal{F}_{t-1}$ represents information available through time $t-1$.

Three Different Mean Quantities

The intercept is $c$, the constant term in the model equation. The unconditional mean is $E[X_t]$, the average level of the process without conditioning on a particular previous observation. Under stationarity, this mean does not depend on $t$; denote it by $\mu$. The conditional mean given the previous value is $E[X_t\mid X_{t-1}]$. It changes with the observed value of $X_{t-1}$.

Taking the expectation of the model equation, using $E[\varepsilon_t]=0$, gives

E[X_t]=c+\phi E[X_{t-1}].

For a stationary process, $E[X_t]=E[X_{t-1}]=\mu$, so the equation becomes $\mu=c+\phi\mu$. Solving yields the stationary mean.

Theorem: For a stationary AR(1) process with $|\phi|<1$ and zero-mean innovations, the unconditional mean is $$ \mu=E[X_t]=\frac{c}{1-\phi}. $$ The intercept $c$ is generally not the process mean. They are equal exactly when $c=0$ or $\phi=0$, equivalently when $c\phi=0$.

The equality condition follows algebraically: $c/(1-\phi)=c$ implies $c=c(1-\phi)$, so $c\phi=0$. Since $|\phi|<1$, the denominator is not zero. Thus either $c=0$, or $\phi=0$, and either condition also makes the two quantities equal. In particular, a zero intercept gives a zero mean for any allowed $\phi$, not just for $\phi=0$.

Under the conditional-mean assumption, the model also gives

E[X_t\mid\mathcal{F}_{t-1}] =c+\phi X_{t-1}.

This is a function of the observed past. It is not usually equal to the unconditional mean. If the previous observation happens to equal $\mu$, then the conditional mean equals $\mu$, because $c+\phi\mu=\mu$. For another previous value, the conditional mean will generally differ.

Why the Stationary-Mean Formula Holds

The derivation uses only the mean equation and the fact that a stationary process has the same mean at consecutive times. The following steps make clear where each assumption enters.

Take expectations.
From $X_t=c+\phi X_{t-1}+\varepsilon_t$, linearity of expectation and $E[\varepsilon_t]=0$ give $E[X_t]=c+\phi E[X_{t-1}]$.

Use stationarity.
Weak stationarity includes a constant mean, so $E[X_t]=E[X_{t-1}]=\mu$. Substitution gives $\mu=c+\phi\mu$.

Rearrange and solve.
Subtract $\phi\mu$ from both sides: $(1-\phi)\mu=c$. Since $|\phi|<1$ implies $1-\phi\ne0$, division gives $\mu=c/(1-\phi)$.

The condition $|\phi|<1$ is the standard causal stationarity condition for this AR(1) model. It also guarantees that the denominator in the mean formula is nonzero. The expectation equation alone can be written for other parameter values if expectations exist, but the stationary causal result presented here should not be applied outside its stated assumptions.

A useful equivalent form centers the process at its mean. Since $c=(1-\phi)\mu$, subtracting $\mu$ from both sides of the model gives

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

This identity shows the role of $c$: it sets the process's long-run level in combination with $\phi$, while $\phi$ describes how a deviation from that level carries forward. The intercept is not itself that level unless the equality condition above holds.

Worked Example: Intercept, Long-Run Mean, and One-Step Mean

A hypothetical monthly index is modeled by $X_t=3+0.5X_{t-1}+\varepsilon_t$. Assume the process is stationary and the innovations are independent and identically distributed as $N(0,4)$, where 4 is the variance. If the latest observed index is $X_{t-1}=8$, find the intercept, unconditional mean, and conditional mean for the next month.

Method. Use $c$ directly for the intercept, $c/(1-\phi)$ for the stationary unconditional mean, and $c+\phi X_{t-1}$ for the conditional mean. The normal innovation distribution has mean zero, as required; its variance does not enter these mean calculations.

Intercept. Comparing the model with $X_t=c+\phi X_{t-1}+\varepsilon_t$, we identify $c=3$ and $\phi=0.5$.

Unconditional mean. The stationarity condition is satisfied because $|0.5|<1$. Therefore

\mu=\frac{3}{1-0.5}=\frac{3}{0.5}=6.

Conditional mean given the latest observation. Substitute $X_{t-1}=8$:

E[X_t\mid X_{t-1}=8]=3+0.5(8)=3+4=7.

The three reported values are: intercept $3$, unconditional mean $6$, and conditional mean $7$. As a check, if the previous observation were the unconditional mean $6$, the conditional mean would be $3+0.5(6)=6$. The method avoids treating the intercept as either kind of mean.

Initial Conditions and the Approach to the Mean

The stationary mean is a long-run property; it need not equal the expected value at every time if the process begins from a nonstationary initial condition. Let $m_t=E[X_t]$. If $E[\varepsilon_t]=0$, then, whether or not the initial value is stationary,

m_t=c+\phi m_{t-1}.

When $|\phi|<1$, the stationary value $\mu=c/(1-\phi)$ solves the same recurrence. Subtracting the equation $\mu=c+\phi\mu$ from the recurrence for $m_t$ gives $m_t-\mu=\phi(m_{t-1}-\mu)$. Repeated substitution yields

m_t-\mu=\phi^t(m_0-\mu), \qquad m_t=\mu+\phi^t(m_0-\mu).

Because $|\phi|<1$, $\phi^t$ tends to zero as $t$ increases. Thus the expected value approaches $\mu$, even when the initial expected value differs from it. This is a statement about the mean recurrence; it does not by itself establish that the full process was stationary from its starting time.

Worked Example: A Nonstationary Starting Mean

A hypothetical quarterly claim-count index follows $X_t=2-0.25X_{t-1}+\varepsilon_t$, where the innovations are independent and identically distributed as $N(0,0.09)$. Start with the deterministic value $X_0=0$. Find the stationary mean and the expected values at times 1 and 2.

Method. First calculate the stationary mean from the model parameters; then apply the expectation recurrence using $m_0=0$. Here $c=2$, $\phi=-0.25$, and the innovations have mean zero and variance $0.09$.

Stationary mean. Since $|-0.25|<1$,

\mu=\frac{2}{1-(-0.25)} =\frac{2}{1.25} =1.6.

Expected value at time 1. With $m_0=0$,

$$ m_1=2-0.25(0)=2. $$

Expected value at time 2. Use the expected value just found:

$$ m_2=2-0.25(2)=2-0.5=1.5. $$

The results $m_1=2$ and $m_2=1.5$ differ from the stationary mean $1.6$, as expected from the nonstationary starting value. Checking the recurrence from the closed form gives $m_1=1.6+(-0.25)(0-1.6)=2$ and $m_2=1.6+(-0.25)^2(0-1.6)=1.5$.

When the Intercept Equals the Mean

A frequent shortcut is to read $c$ as the process's average level. The shortcut is generally wrong because the previous observation contributes $\phi X_{t-1}$. The correct comparison is

\mu-c=\frac{c}{1-\phi}-c =\frac{c\phi}{1-\phi}.

Thus the intercept and unconditional mean are equal if and only if $c\phi=0$, which means $c=0$ or $\phi=0$. For example, if $c=0$ and $\phi=0.6$, then $\mu=0/(1-0.6)=0$, so the intercept and mean agree even though $\phi\ne0$. If instead $c\ne0$ and $\phi\ne0$, they differ.

A second common confusion is to call $c+\phi X_{t-1}$ the unconditional mean. It is the conditional mean given the previous information. The unconditional mean averages over the process's possible previous values; in the stationary model, that averaging leads to $c+\phi\mu=\mu$.

Worked Example: Zero Intercept Does Not Require Zero Dependence

A hypothetical weekly demand deviation is modeled as $X_t=0+0.6X_{t-1}+\varepsilon_t$, with independent and identically distributed innovations $N(0,1)$. Determine the stationary unconditional mean and the conditional mean when the previous deviation is 5. Explain whether the intercept equals the unconditional mean.

Method. Apply the stationary mean formula, then evaluate the conditional mean at the specified previous value. The innovation distribution has mean zero and variance 1, and $|0.6|<1$.

Unconditional mean.

\mu=\frac{0}{1-0.6}=\frac{0}{0.4}=0.

Conditional mean.

E[X_t\mid X_{t-1}=5]=0+0.6(5)=3.

The intercept and unconditional mean are both zero because $c=0$. The conditional mean is 3 for the specified previous observation. These values are consistent: the conditional mean depends on the current information, whereas the unconditional mean describes the stationary average across time. This example avoids the incorrect claim that equality between intercept and mean requires $\phi=0$.

Common Mistakes and Practical Checks

Using the intercept as the unconditional mean. The incorrect calculation is $\mu=c$. For $c=3$ and $\phi=0.5$, this gives $3$, but the correct calculation is $\mu=3/(1-0.5)=6$.
Giving an incomplete equality condition. It is incorrect to say that $c$ equals $\mu$ only when $\phi=0$. Since $c/(1-\phi)=c$ exactly when $c\phi=0$, equality also holds whenever $c=0$.
Confusing a conditional mean with the unconditional mean. Evaluate $c+\phi X_{t-1}$ only when a previous value or past information is specified. For the stationary unconditional mean, solve $\mu=c+\phi\mu$.
Ignoring the innovation mean. The formula $\mu=c/(1-\phi)$ assumes $E[\varepsilon_t]=0$. If instead the innovation mean is $a$, the expectation equation is $\mu=c+\phi\mu+a$, giving $\mu=(c+a)/(1-\phi)$, provided the stationary assumptions hold.
Using the stationary result outside its stated model conditions. Check the assumptions and $|\phi|<1$ before interpreting $c/(1-\phi)$ as the mean of a stationary causal AR(1) process.

Key takeaway: For a stationary AR(1) process with zero-mean innovations, the unconditional mean is $c/(1-\phi)$. The one-step conditional mean is $c+\phi X_{t-1}$. The intercept equals the unconditional mean exactly when $c=0$ or $\phi=0$.

Check Your Understanding

For each model, identify the requested mean and check whether the stationary condition is satisfied.

A model has $c=4$, $\phi=0.25$, and zero-mean innovations. Find the stationary unconditional mean and the conditional mean when $X_{t-1}=8$. (Answer: unconditional mean $16/3\approx5.3333$; conditional mean $6$.)
A stationary model has $c=0$ and $\phi=-0.4$. Find the unconditional mean and state whether it equals the intercept. (Answer: mean $0$; yes, because $c=0$.)
A model has $c=1.5$, $\phi=0.5$, and zero-mean innovations. Find the stationary mean and the conditional mean given $X_{t-1}=2.4$. (Answer: unconditional mean $3$; conditional mean $2.7$.)
A stationary model has $c=-2$, $\phi=-0.5$, and zero-mean innovations. Find its unconditional mean and the conditional mean given $X_{t-1}=-4$. (Answer: unconditional mean $-4/3$; conditional mean $0$.)
For $X_t=3+0.5X_{t-1}+\varepsilon_t$, suppose $E[X_0]=1$ and the innovations have mean zero. Find $E[X_1]$, $E[X_2]$, and the stationary mean. (Answer: $E[X_1]=3.5$, $E[X_2]=4.75$, stationary mean $6$.)

What You'll Learn