Moving Average Models

A formal moving-average model is not the same thing as a rolling average smoother. A rolling average replaces an observation with an average of neighboring observations. An MA model describes the series as a weighted sum of current and past innovations, which are the unpredictable shocks left after forecasting.

An MA$(q)$ model is

$$y_t=c+{\epsilon }_t+{\theta }_1{\epsilon }_{t-1}+\cdots +{\theta }_q{\epsilon }_{t-q}.$$

The current value is affected by a new shock ${\epsilon }_t$ and by a finite number of older shocks. After $q$ lags, those shocks no longer directly enter the equation. That finite shock memory gives MA models their diagnostic signature: a pure MA$(q)$ process has an ACF that cuts off after lag $q$, while the PACF decays more gradually.

The intuition is easier if $y_t$ is a residual process. Suppose a demand baseline misses an unexpected promotion spike. A moving-average error term lets that surprise affect the next few residuals while the system settles back down. The model does not say past observed demand directly causes current demand; it says past forecast errors carry information about unobserved disturbances.

MA terms appear inside ARMA, ARIMA, and SARIMA. They are estimated indirectly because the innovations are not observed before fitting. This makes invertibility important: without it, different MA parameter values can imply the same autocorrelation behavior. In practice, MA order should be judged by residual diagnostics and forecast error metrics, not by smoothing appearance.

Connections

Moving-average error terms are the MA side of ARMA and ARIMA. Their diagnostic footprint appears in autocorrelation and partial autocorrelation, while simple rolling averages belong more naturally with statistical forecasting baselines.

References

• Hyndman & Athanasopoulos, FPP3: Moving average models
• statsmodels ARIMA API

Nav

Section — Time-Series Forecasting

← Autoregressive Models ARMA →