Specifying an AR(1) Model

What an AR(1) Specification Says

A time series records a variable in time order. An autoregressive model uses earlier observations of that variable to describe its current value. The simplest such model uses just the immediately preceding observation: it is an autoregressive model of order one, or AR(1). Specifying the model means stating how the current value depends on that lagged value, the intercept, and a new random innovation.

For example, a hypothetical insurer might track a monthly index of claim-processing delays. A high value in one month may be associated with a higher or lower value in the next month, while new operational events create unpredictable variation. The model describes this time-ordered dependence; by itself, it does not establish a causal explanation for it.

Definition: An AR(1) model is $$ Y_t=c+\phi Y_{t-1}+\varepsilon_t, $$ where $Y_t$ is the observation at time $t$, $c$ is a constant intercept, $\phi$ is the autoregressive coefficient, and $\varepsilon_t$ is the innovation at time $t$. The innovation is white noise if $E[\varepsilon_t]=0$, $\operatorname{Var}(\varepsilon_t)=\sigma_\varepsilon^2$ for every $t$, and $\operatorname{Cov}(\varepsilon_t,\varepsilon_s)=0$ for $t\ne s$. White noise is uncorrelated over time; it is not necessarily independent.

The equation is a specification, not a claim that every series follows this model. It says that the previous observation enters linearly, with coefficient $\phi$, and that the remainder is represented by an innovation with mean zero and constant variance. A particular application may assume more about the innovations—for example, that they are independent and normally distributed—but those properties do not follow just from the definition of white noise.

To calculate a sequence from the model, one also needs an initial value such as $Y_0$ and the relevant innovation values. The initial value is the starting observation, not an additional parameter in the AR(1) equation. In applications, parameters are usually estimated from observed data; here, worked examples treat them as given so that the model's components are clear.

Interpreting the Three Components

The intercept $c$ is the constant part of the model's one-step calculation. The coefficient $\phi$ multiplies the previous observation: holding the innovation fixed, a one-unit increase in $Y_{t-1}$ changes the model's current value by $\phi$ units. The innovation $\varepsilon_t$ adds the new, time-specific departure from the part explained by the intercept and lagged observation.

The sign of $\phi$ determines the direction of the lagged contribution. If $\phi$ is positive, a positive previous value contributes positively to the current value, and a negative previous value contributes negatively. If $\phi$ is negative, the lagged contribution has the opposite sign. If $\phi=0$, the previous observation makes no direct contribution to the current equation. These interpretations concern the equation itself; they do not, on their own, establish whether the process is stationary or describe its full long-run behavior.

Worked Example: A Positive Lag-One Coefficient

A hypothetical monthly delay index follows $Y_t=1+0.5Y_{t-1}+\varepsilon_t$. Assume $\{\varepsilon_t\}$ are independent with distribution $N(0,0.25)$, where the second parameter is the variance. Starting at $Y_0=2$, the observed innovations for the next three months are $0.2$, $-0.4$, and $0.1$. Calculate $Y_1,Y_2,Y_3$, and interpret the coefficient. Use the specified innovations as realizations; the distribution describes how innovations are modeled, not the values they must take.

Method. Substitute the previous observation and that period's innovation into the AR(1) equation at each time step.

At $t=1$: $$ Y_1=1+0.5(2)+0.2=1+1+0.2=2.2. $$

At $t=2$: Use the calculated $Y_1$, not the initial value: $$ Y_2=1+0.5(2.2)-0.4=1+1.1-0.4=1.7. $$

At $t=3$: $$ Y_3=1+0.5(1.7)+0.1=1+0.85+0.1=1.95. $$

The lag coefficient is positive: with the innovation held fixed, increasing the previous index by one unit increases the current model value by $0.5$ units. The calculations check against the recursion: each value uses the immediately preceding calculated value and the correct month's innovation. A common mistake is to keep using $Y_0$ at every step.

Worked Example: A Negative Lag-One Coefficient

A hypothetical monthly workload deviation is modeled by $Y_t=0.4-0.6Y_{t-1}+\varepsilon_t$. Assume the innovations are independent with distribution $N(0,0.09)$. Given $Y_0=1.5$, the next three observed innovations are $0.1$, $0$, and $-0.2$. Calculate the three subsequent values and explain how the negative coefficient affects the first step.

Method. Apply the specified recursion in order; the negative sign belongs to the coefficient multiplying the lagged observation.

At $t=1$: $$ Y_1=0.4-0.6(1.5)+0.1 =0.4-0.9+0.1=-0.4. $$

At $t=2$: $$ Y_2=0.4-0.6(-0.4)+0 =0.4+0.24=0.64. $$

At $t=3$: $$ Y_3=0.4-0.6(0.64)-0.2 =0.4-0.384-0.2=-0.184. $$

At the first step, the positive lagged value $1.5$ contributes $-0.9$, because its coefficient is $-0.6$. At the second step, the lagged value is negative, so multiplying it by $-0.6$ gives a positive contribution of $0.24$. The arithmetic checks by substituting each result into the next step. The main trap is to treat a negative coefficient as a negative contribution regardless of the sign of the lagged value; the sign of the product depends on both factors.

Worked Example: A Zero Lag-One Coefficient

A hypothetical monthly service index has the specification $Y_t=3+0Y_{t-1}+\varepsilon_t$. Assume the innovations are independent with distribution $N(0,1)$. Starting from $Y_0=8$, suppose the next two innovations are $0.5$ and $-1$. Calculate the two subsequent values and state what role $Y_0$ plays in these calculations.

Method. Substitute $\phi=0$ into the equation. The lagged-value term is zero at each time.

At $t=1$: $$ Y_1=3+0(8)+0.5=3.5. $$

At $t=2$: $$ Y_2=3+0(3.5)-1=2. $$

The starting value $Y_0=8$ has no direct effect on $Y_1$, because its coefficient is zero. Likewise, $Y_1$ has no direct effect on $Y_2$. The results check directly: each is the intercept plus that period's innovation. A common mistake is to carry the large initial value forward even though the model assigns its lagged contribution a coefficient of zero.

Why the Coefficient Has This Interpretation

The interpretation of $\phi$ follows directly by comparing two possible previous values under the same intercept and innovation. Suppose one calculation uses previous value $y$, and another uses $y+\Delta$, where $\Delta$ is a change in the previous observation. Let their resulting current values be $Y_t^{(1)}$ and $Y_t^{(2)}$, with the same $\varepsilon_t$. The model gives

\begin{aligned} Y_t^{(1)}&=c+\phi y+\varepsilon_t,\\ Y_t^{(2)}&=c+\phi(y+\Delta)+\varepsilon_t. \end{aligned}

Subtracting the first equation from the second cancels the shared intercept and innovation:

Y_t^{(2)}-Y_t^{(1)} =\phi(y+\Delta-y) =\phi\Delta.

Thus, for a one-unit change ($\Delta=1$), the model's current value changes by $\phi$ units when the innovation is held fixed. This is an algebraic interpretation of the coefficient, not a claim that changing an observed value would causally change a real-world outcome.

There is also a prediction interpretation under an additional condition. If the current innovation has conditional mean zero given the information available through time $t-1$, then taking the conditional expectation of the model equation gives

E[Y_t\mid Y_{t-1},Y_{t-2},\ldots] =c+\phi Y_{t-1}.

This condition holds, for example, when innovations are independent over time and independent of the past. Merely knowing that innovations are uncorrelated with one another does not in every model establish this conditional-mean property. When it does hold, the expression is the model's one-step conditional prediction; the realized observation also includes $\varepsilon_t$.

Reading an AR(1) Equation Carefully

The intercept and coefficient have distinct roles. The intercept is the additive constant in each one-step equation; it is not automatically the overall or long-run average level of the series. The coefficient describes the effect of the immediately preceding value in that equation. To see why it is important not to confuse these roles, consider $Y_t=2+0.5Y_{t-1}+\varepsilon_t$. If $Y_{t-1}=6$ and the current innovation is $0.2$, the model gives

$$ Y_t=2+0.5(6)+0.2=2+3+0.2=5.2. $$

Setting the lagged value to zero would produce $2.2$, not the value implied by the stated previous observation. The intercept alone is not the one-step value unless the lagged contribution is zero.

For an observed time series, the innovation at a given time is often treated as the difference between the observed value and the value predicted by the intercept and lagged observation:

\varepsilon_t=Y_t-(c+\phi Y_{t-1}).

This is the model's residual-like, unexplained component at time $t$. A proposed AR(1) specification is useful only if its assumptions about these innovations are reasonable for the application. In particular, systematic patterns left in the innovations would indicate that the model has not captured all the time dependence represented in the data.

The coefficient's sign is not enough to establish whether a fitted model is appropriate, and the equation alone does not assert that its process is stationary. The examples above demonstrate finite-step calculations for specified coefficients and innovations; they make no stationarity claim. In model interpretation, keep separate the equation's immediate arithmetic, the stochastic assumptions on innovations, and any additional conditions required for properties of the full process.

Key takeaway: In $Y_t=c+\phi Y_{t-1}+\varepsilon_t$, $c$ is the additive intercept, $\phi$ scales the previous observation, and $\varepsilon_t$ is the new innovation. The sign and size of $\phi$ determine the direction and amount of the lagged contribution; the innovation assumptions determine what is meant by the unpredictable part.

Check Your Understanding

For each question, use the stated equation and innovation values. Treat the given parameters as specified rather than estimated.

A hypothetical monthly inventory deviation follows $Y_t=0.5+0.4Y_{t-1}+\varepsilon_t$. The innovations are independent with distribution $N(0,0.16)$, $Y_0=5$, and $\varepsilon_1=-0.2,\varepsilon_2=0.3$. Find $Y_1$ and $Y_2$. (Answer: $Y_1=2.3,\;Y_2=1.72$.)
A hypothetical service index follows $Y_t=1-0.5Y_{t-1}+\varepsilon_t$, with independent $N(0,0.09)$ innovations. If $Y_0=2$, $\varepsilon_1=0.1$, and $\varepsilon_2=-0.2$, calculate $Y_1$ and $Y_2$. (Answer: $Y_1=0.1,\;Y_2=0.75$.)
A hypothetical monthly adjustment is specified by $Y_t=-1+0Y_{t-1}+\varepsilon_t$, with independent $N(0,0.25)$ innovations. Starting from $Y_0=7$, use $\varepsilon_1=0.25$ and $\varepsilon_2=-0.5$ to calculate both new values, and state whether $Y_0$ affects $Y_1$. (Answer: $Y_1=-0.75,\;Y_2=-1.5$; $Y_0$ has no direct effect.)
A hypothetical monthly workload follows $Y_t=2+0.7Y_{t-1}+\varepsilon_t$, with independent $N(0,0.36)$ innovations. If $Y_t=3$ and the next observed innovation is $-0.4$, calculate the one-step predicted value before the innovation and the realized next value after it. (Answer: predicted value $4.1$; realized value $3.7$.)

What You'll Learn