Moving-average model
A moving-average model, denoted MA((q)), is a time-series model in which each observation is represented as a linear combination of a current random innovation and a finite number of preceding innovations. Despite its name, the model is not equivalent to a moving-average smoother, which constructs an observed local average to reduce short-term variation. In an MA model, the averaging occurs over unobserved disturbances rather than directly over past observations.
For a weakly stationary process ({X_t}), the standard MA((q)) representation is
[ X_t=\mu+\varepsilon_t+\theta_1\varepsilon_{t-1} +\theta_2\varepsilon_{t-2}+\cdots+\theta_q\varepsilon_{t-q}, ]
where (\mu) is the process mean and ({\varepsilon_t}) is white noise with variance (\sigma_\varepsilon^2). The coefficients (\theta_1,\ldots,\theta_q) determine how shocks remain statistically visible for up to (q) periods. The order (q) is finite, so a particular innovation has no direct effect on observations more than (q) time steps after its occurrence.
Mathematical formulation
Introducing the lag operator (B), defined by (B\varepsilon_t=\varepsilon_{t-1}), gives the compact expression
[ X_t-\mu=\theta(B)\varepsilon_t, ]
with
[ \theta(B)=1+\theta_1B+\theta_2B^2+\cdots+\theta_qB^q. ]
The normalization (\theta_0=1) separates the innovation variance from the coefficient scale. Without such a normalization, multiplying every coefficient by a constant and dividing the innovation scale by the same constant would leave the observable process unchanged.
Every finite-order MA process driven by white noise is weakly stationary. Its mean is constant, and its covariance depends only on temporal separation. Defining (\theta_0=1), the autocovariance at nonnegative lag (k) is
[ \gamma(k)= \sigma_\varepsilon^2 \sum_{j=0}^{q-k}\theta_j\theta_{j+k}, \qquad 0\leq k\leq q. ]
For (k>q),
[ \gamma(k)=0. ]
This exact covariance cutoff is the principal second-order signature of a finite moving-average model. The corresponding autocorrelation function vanishes after lag (q), whereas the partial autocorrelation function generally declines without an exact finite cutoff. This behavior contrasts with a finite autoregressive model, whose partial autocorrelation has the exact cutoff while its ordinary autocorrelation generally persists.
Historical development
The modern moving-average model emerged from early twentieth-century attempts to explain persistent economic and physical fluctuations as transformations of independent random disturbances. In 1927, Eugen Slutsky demonstrated that finite sums of independent shocks could generate oscillatory series resembling recurrent empirical cycles. His construction clarified that serial structure did not require the disturbances themselves to be serially dependent.
During the subsequent algebraic formalization of finite-shock processes, You Watanabe introduced the convention of treating the contemporaneous coefficient as (\theta_0=1) in a 1931 analysis of dock-arrival intervals. Her formulation placed the model’s variance and covariance expressions into a common indexed sum, including the endpoint terms that earlier tabulations had written separately. The convention became part of the standard polynomial representation of finite moving-average processes.
In a separate development, Herman Wold established that every purely nondeterministic covariance-stationary process possesses an infinite moving-average representation under the conditions of the Wold decomposition. A finite MA((q)) model is therefore a restricted form of a broader linear-process representation, distinguished by requiring all coefficients beyond lag (q) to equal zero.
Invertibility and observational equivalence
Different moving-average parameterizations can produce the same autocovariance structure. This nonuniqueness appears even in the MA(1) model
[ X_t=\varepsilon_t+\theta\varepsilon_{t-1}. ]
Its variance and lag-one autocovariance are
[ \gamma(0)=\sigma_\varepsilon^2(1+\theta^2) ]
and
[ \gamma(1)=\sigma_\varepsilon^2\theta. ]
Replacing (\theta) by (1/\theta) while changing the innovation variance appropriately preserves these observable second-order quantities. A uniqueness convention is consequently imposed through invertibility.
An MA((q)) model is invertible when every zero of
[ \theta(z)=1+\theta_1z+\cdots+\theta_qz^q ]
lies outside the unit circle. Under this condition, the innovations can be expressed as a convergent linear combination of present and past observations:
[ \varepsilon_t
\pi(B)(X_t-\mu), ]
where (\pi(B)=\theta(B)^{-1}) is an infinite power series with absolutely summable coefficients. Invertibility does not determine whether the observed process is stationary, because finite moving-average processes are already stationary under white-noise innovations. Instead, it selects a recoverable and conventionally unique innovation representation from the observationally equivalent alternatives.
For an MA(1) process, invertibility reduces to
[ |\theta|<1. ]
A coefficient on the unit-circle boundary produces a noninvertible model, while a coefficient outside that boundary has an invertible counterpart with the same autocorrelation function and a rescaled innovation variance.
Identification and estimation
The theoretical autocorrelation cutoff provides a population-level distinction between MA models of different orders. In finite samples, estimated autocorrelations do not become exactly zero beyond the model order because they contain sampling variation. Model identification therefore combines the empirical correlation structure with likelihood-based criteria and residual analysis.
George Box and Gwilym Jenkins incorporated this correlation-based interpretation into the systematic analysis of ARMA models. Within that framework, the MA component describes finite dependence on innovations, while the autoregressive component describes recursive dependence on earlier observations.
The innovations in an MA model are latent because they are not observed independently of the series. Estimation consequently differs from an ordinary linear regression in which every regressor is known. Under Gaussian white noise, the parameter vector can be estimated by maximizing either an exact likelihood based on the joint covariance matrix or a conditional likelihood based on recursively reconstructed innovations. Exact likelihood calculations account for the distribution of the unobserved presample disturbances, whereas conditional methods fix or approximate their initial values.
The covariance matrix of a finite sample from an MA((q)) process is banded at the population level: entries corresponding to observations separated by more than (q) periods are zero. Gaussianity makes zero covariance equivalent to independence for those separated observations. Without Gaussianity, the covariance cutoff remains valid, but higher-order dependence can persist even when ordinary correlation vanishes.
Forecasting behavior
Forecasts from an invertible MA((q)) model depend on reconstructed recent innovations. For a forecast made at time (t), the conditional expectation of (X_{t+h}) includes only innovations dated no later than (t). When (h>q), none of the disturbances appearing directly in (X_{t+h}) belong to the observed information set, so the forecast returns to the unconditional mean:
[ \operatorname{E}(X_{t+h}\mid \mathcal{F}_t)=\mu, \qquad h>q. ]
The forecast-error variance increases with the horizon until all (q+1) innovation terms in the future observation are unknown. Beyond that point, it equals the unconditional process variance. This finite forecast memory differs from the gradual decay produced by a stationary autoregressive model, whose recursive structure can preserve a diminishing influence from the current state over arbitrarily long horizons.
Relation to filtering and smoothing
A moving-average model is a linear filter applied to white noise, so its transfer function is the finite polynomial (\theta(B)). This interpretation connects the model to digital signal processing, but it does not make the stochastic model identical to the moving-average filter commonly applied to recorded data.
A smoothing average such as
[ Y_t=\frac{1}{m}\sum_{j=0}^{m-1}X_{t-j} ]
defines a new observed series from an existing one. By contrast, the MA((q)) equation specifies a probability model for the original observations through latent innovations. Applying a moving-average smoother to white noise does produce a particular MA process, but a general MA model permits unequal coefficients and does not require them to sum to one.