Autoregressive Models (AR)
What you'll learn: How autoregressive models predict future values by using a linear combination of past observations plus random noise.
The Core Idea
An autoregressive model assumes that the current value in a time series depends on its own previous values. Think of it like saying: "Tomorrow's temperature is mostly based on today's temperature, yesterday's temperature, and maybe a few days before that—plus some randomness."
Mathematically, an AR model of order p (written as AR(p)) looks like:
y_t = c + φ₁·y_{t-1} + φ₂·y_{t-2} + ... + φₚ·y_{t-p} + ε_t
Where:
y_tis the current valuecis a constant (intercept)φ₁, φ₂, ..., φₚare learned coefficients (weights on past values)ε_tis white noise (random error)pis the number of lag terms used
How It Works
The model "regresses" the variable on itself from previous time steps—hence "autoregressive." The order p tells you how many past steps matter. An AR(1) model only looks one step back, while AR(3) considers the three most recent observations.
The coefficients determine how much influence each past value has. If φ₁ = 0.7, today's value has 70% weight on yesterday's value. These coefficients are learned from your training data using methods like least squares.
Connection to Previous Concepts
Unlike Moving Average (MA) models that use past errors, AR models use past values directly. Both rely on stationarity and autocorrelation patterns you've already explored. You'll use lag features and autocorrelation plots to choose the right order p.
Key Takeaway: Autoregressive models predict future values as weighted sums of past values plus noise, making them powerful for capturing temporal dependencies in stationary time series.