Moving Average Models (MA)
What you'll learn: How a moving average model represents a time series as a weighted sum of recent random shocks, making it powerful for capturing short-term dependencies.
The Core Idea
A Moving Average (MA) model doesn't average past observations like a rolling mean. Instead, it models your time series as a linear combination of white noise error terms from previous time steps. Think of it as your series "remembering" recent unpredictable shocks that pushed it off course.
The MA(q) model looks like this:
y_t = μ + ε_t + θ₁ε_{t-1} + θ₂ε_{t-2} + ... + θ_qε_{t-q}
Where:
y_tis your observation at time tμis the mean levelε_tis white noise (random shock) at time tθ₁, θ₂, ..., θ_qare weights (coefficients)qis the order – how many past errors you include
The Intuition
Imagine stock prices reacting to news. An MA(2) model says: "Today's price equals the baseline plus today's random news plus yesterday's lingering impact plus the day-before's fading impact." Each shock ripples forward but eventually fades out.
Why MA Matters
Unlike autoregressive models (which we build toward), MA models capture how recent unexpected events influence current values. The autocorrelation plot cuts off sharply after lag q, which helps you identify the right order. MA processes are always stationary (assuming white noise is), making them mathematically tractable.
Key Takeaway: Moving Average models express time series as weighted combinations of white noise shocks, capturing how random disturbances propagate through time with diminishing influence.