What an MA(1) Model Specifies
An autoregressive model describes the current value using earlier values of the series. A moving-average model instead describes it using current and past innovations: the unpredictable shocks in a time series. For an MA(1), the current observation depends on the current innovation and exactly one lagged innovation. Specifying the model means stating its mean, its coefficient on that lagged shock, and the properties of the innovation sequence.
The subscript \(t-1\) means the innovation from the immediately preceding time period, not the previous observation \(X_{t-1}\). That distinction is fundamental: in an MA(1), the lagged term is \(\varepsilon_{t-1}\). The model does not say that \(X_t\) is a weighted average of \(X_t\) and \(X_{t-1}\).
White noise requires zero mean, constant variance, and no correlation across distinct times. Those conditions alone do not imply independence or normality. If a problem specifies independent normal innovations, that is a stronger assumption, and it should be stated separately. The parameter \(\sigma_\varepsilon^2\) is the variance of each innovation; \(\sigma_\varepsilon\) is its standard deviation. Do not confuse the two.
Interpreting the Parameters
The constant \(\mu\) is the baseline level of the series. The current innovation enters with coefficient 1, while \(\theta\) determines how much of the preceding innovation enters the current value. A positive \(\theta\) gives the lagged innovation the same sign as its realization; a negative \(\theta\) reverses its sign. The size \(|\theta|\) describes the scale of that lagged contribution, measured in the same units as \(X_t\) per unit of innovation.
The coefficient is not itself a correlation, and it is not the innovation variance. For instance, \(\theta=0.5\) means that a preceding innovation of 2 contributes \(0.5(2)=1\) to the current value. Separately, \(\sigma_\varepsilon^2\) describes the variability of the innovation sequence. Specifying one does not determine the other.
Some texts use the form \(X_t=\mu+\varepsilon_t-\alpha\varepsilon_{t-1}\). This is the same sign convention after setting \(\theta=-\alpha\). Always read the equation before interpreting a reported coefficient: a positive number attached to a minus sign corresponds to a negative lag coefficient under the convention used here.
Worked Example: Calculate a Value from Two Innovations
A hypothetical daily index is specified by \(X_t=8+\varepsilon_t+0.5\varepsilon_{t-1}\). Assume the innovations are independent and each has distribution \(N(0,1)\), so the innovation mean is 0 and its variance is 1. On two consecutive days, the realized innovations are \(\varepsilon_6=2\) and \(\varepsilon_7=-4\). Calculate \(X_7\), and identify the contribution from the lagged innovation.
Method. Substitute the time-7 innovation and the time-6 innovation into the MA(1) equation. The mean is 8, the coefficient is \(\theta=0.5\), and the lagged innovation is \(\varepsilon_{7-1}=\varepsilon_6\).
Current-innovation contribution: \(\varepsilon_7=-4\).
Lagged-innovation contribution: \(\theta\varepsilon_6=(0.5)(2)=1\).
Combine the terms: \(X_7=8+(-4)+1=5\).
Therefore \(X_7=5\), and the lagged innovation contributes \(1\). The calculation checks because the mean and two innovation contributions add to \(8-4+1=5\). The common mistake to avoid is reporting 3 after substitution: that value does not equal the sum of the stated terms.
Why the Lagged Term Has This Interpretation
The MA(1) equation separates the current value into a baseline and two shock contributions. To make the “current innovation” interpretation precise, suppose the innovations have conditional mean zero given all earlier innovations. This property holds, for example, when the innovations are independent with mean zero. Let \(\mathcal{H}_{t-1}\) denote the information consisting of the innovations through time \(t-1\). In the model, \(\varepsilon_{t-1}\) is known given \(\mathcal{H}_{t-1}\), while the conditional mean of \(\varepsilon_t\) is zero. Therefore:
Subtracting this conditional expectation from the current value gives \(X_t-E[X_t\mid\mathcal{H}_{t-1}]=\varepsilon_t\). Thus, under the stated conditional-mean assumption, the current innovation is the new unpredictable part of \(X_t\) after accounting for the baseline and the known lagged innovation. White noise by itself guarantees zero covariance across times, but not this conditional-mean statement in every possible distributional setup; that is why the stronger condition was stated before using it.
This also clarifies the limited reach of a lagged innovation. In the equation for \(X_t\), \(\varepsilon_{t-1}\) enters with coefficient \(\theta\). It is not a claim that the earlier observation \(X_{t-1}\) has coefficient \(\theta\), since \(X_{t-1}\) contains its own innovation terms and mean.
Worked Example: Interpret a Negative MA(1) Coefficient
A hypothetical monthly claims index follows \(X_t=120+\varepsilon_t-0.4\varepsilon_{t-1}\). Assume the innovations are independent and distributed as \(N(0,9)\), with mean 0 and variance 9. For one month, \(\varepsilon_t=6\) and \(\varepsilon_{t-1}=-5\). State the MA(1) coefficient under the plus-sign convention, calculate both innovation contributions, and find \(X_t\).
Method. Rewrite the model in the course convention \(X_t=\mu+\varepsilon_t+\theta\varepsilon_{t-1}\). Comparing coefficients gives \(\mu=120\) and \(\theta=-0.4\). The innovation variance is 9, and its standard deviation is \(\sqrt{9}=3\); neither value replaces the coefficient.
Current contribution: \(\varepsilon_t=6\).
Lagged contribution: \(\theta\varepsilon_{t-1}=(-0.4)(-5)=2\). Both factors are negative, so their product is positive.
Calculate the index: \(X_t=120+6+2=128\).
The result is \(X_t=128\), with contributions of 6 from the current innovation and 2 from the lagged innovation. The common sign trap is to say the lagged contribution is \(-2\) merely because the equation displays a minus sign. The lagged innovation itself is \(-5\), so the product \((-0.4)(-5)\) is positive.
Specifying a Model Clearly
A complete MA(1) specification should make clear what is random and what is fixed. State the equation, identify the constant mean and lag coefficient, and give the innovation assumptions and variance. If normality or independence is being assumed, say so explicitly rather than treating either as part of the definition of white noise.
Use \(\varepsilon_t\) for the current innovation and \(\varepsilon_{t-1}\) for the immediately preceding innovation.
In \(X_t=\mu+\varepsilon_t+\theta\varepsilon_{t-1}\), name \(\mu\), \(\theta\), and \(\sigma_\varepsilon^2\) separately.
Give the zero mean, constant variance, and zero covariance across distinct times; add independence or a distribution only when specified.
Multiply the lagged innovation by \(\theta\), retain its sign, and add it to the mean and current innovation.
Worked Example: Translate a Minus-Sign Specification
A hypothetical monthly service-volume index is written as \(X_t=25+\varepsilon_t-0.25\varepsilon_{t-1}\). Assume independent innovations with distribution \(N(0,4)\), so their mean is 0 and variance is 4. For a particular month, \(\varepsilon_t=-1\) and \(\varepsilon_{t-1}=3\). Express the model using the course’s plus-sign convention, identify its parameters, and calculate the index.
Method. Match the lag term \(-0.25\varepsilon_{t-1}\) to \(+\theta\varepsilon_{t-1}\). Hence \(\theta=-0.25\), while \(\mu=25\) and \(\sigma_\varepsilon^2=4\). The equivalent specification is \(X_t=25+\varepsilon_t+(-0.25)\varepsilon_{t-1}\).
Current contribution: \(\varepsilon_t=-1\).
Lagged contribution: \(\theta\varepsilon_{t-1}=(-0.25)(3)=-0.75\).
Calculate the index: \(X_t=25+(-1)+(-0.75)=23.25\).
Thus \(X_t=23.25\). The result checks because the lagged contribution is negative when a positive preceding innovation is multiplied by a negative coefficient. A common error is to read the displayed \(0.25\) as \(\theta=+0.25\) without accounting for the minus sign in the equation.
Common Specification Traps
- Using the previous observation instead of the previous innovation. The model contains \(\varepsilon_{t-1}\), not \(X_{t-1}\).
- Dropping the sign. Calculate \(\theta\varepsilon_{t-1}\) as a product; do not infer its sign from \(\theta\) alone.
- Confusing variance with standard deviation. If \(\varepsilon_t\sim N(0,9)\), its variance is 9 and its standard deviation is 3.
- Calling the coefficient a correlation or variance. The coefficient \(\theta\) scales the lagged innovation’s contribution; the innovation variance is a distinct parameter.
- Assuming more than white noise states. Zero covariance across time does not, by itself, assert independence or normality.
A clear specification prevents these errors before any further analysis. It also keeps parameter interpretation separate from calculations involving particular realized innovations. The focus here is the model equation and its terms; properties such as the process autocovariances require additional steps.
Check Your Understanding
For each problem, use \(X_t=\mu+\varepsilon_t+\theta\varepsilon_{t-1}\) unless another form is stated. Keep the innovation variance distinct from the MA(1) coefficient.
- A hypothetical weekly index has \(X_t=14+\varepsilon_t+0.2\varepsilon_{t-1}\), with independent \(N(0,4)\) innovations. If \(\varepsilon_t=3\) and \(\varepsilon_{t-1}=-5\), find the lagged contribution and \(X_t\). (Answer: lagged contribution \(=-1\); \(X_t=16\).)
- A process is written \(X_t=50+\varepsilon_t-0.6\varepsilon_{t-1}\), with white-noise innovations of variance 16. Give \(\theta\) under the plus-sign convention and give the innovation standard deviation. (Answer: \(\theta=-0.6\); standard deviation \(=4\).)
- A hypothetical index has mean 30 and lag coefficient \(\theta=-0.3\). Its independent innovations have distribution \(N(0,1)\). If the current innovation is 2 and the preceding innovation is 4, calculate \(X_t\) and identify the two shock contributions. (Answer: current contribution \(=2\), lagged contribution \(=-1.2\), and \(X_t=30.8\).)
- A model is specified by \(X_t=75+\varepsilon_t+0.5\varepsilon_{t-1}\), where the innovations are independent with mean 0 and variance 25. For \(\varepsilon_t=-3\) and \(\varepsilon_{t-1}=6\), calculate \(X_t\), and state the innovation standard deviation. (Answer: \(X_t=75-3+3=75\); standard deviation \(=5\).)
- A hypothetical daily series is given as \(X_t=9+\varepsilon_t-0.4\varepsilon_{t-1}\). The independent innovations have distribution \(N(0,9)\), and \(\varepsilon_t=1\), \(\varepsilon_{t-1}=-2\). Rewrite the lag coefficient under the plus-sign convention, calculate \(X_t\), and state the innovation variance. (Answer: \(\theta=-0.4\); lagged contribution \(=0.8\); \(X_t=10.8\); innovation variance \(=9\).)