ARIMA Models
What you'll learn: How ARIMA extends ARMA models to forecast real-world, non-stationary time series by incorporating differencing.
From ARMA to ARIMA
You've learned that ARMA models work beautifully on stationary data—but most real-world time series aren't stationary. They have trends, seasonal patterns, or changing means. ARIMA (AutoRegressive Integrated Moving Average) solves this by adding an integration step to handle non-stationarity.
The "I" in ARIMA stands for "Integrated," which refers to differencing the series to make it stationary before applying ARMA. Once you've differenced the data d times to achieve stationarity, you fit an ARMA(p,q) model to the transformed series. Forecasts are then "integrated" back (undifferenced) to the original scale.
ARIMA(p,d,q) Parameters
An ARIMA model has three key parameters:
- p: Order of the autoregressive (AR) component—how many past values influence the current value
- d: Degree of differencing—how many times you difference the series to make it stationary
- q: Order of the moving average (MA) component—how many past forecast errors influence the current value
Think of it like this: If your sales data shows an upward trend, raw ARMA won't work. But difference the series once (d=1) to remove the trend, fit ARMA(p,q) to the differences, then add back the trend structure when forecasting—that's ARIMA(p,1,q).
Parameter Selection
Choose d by testing for stationarity after each differencing step. Select p and q using ACF/PACF plots (covered earlier) or information criteria like AIC/BIC, which balance model fit against complexity.
Key Takeaway: ARIMA extends ARMA to non-stationary data by differencing (d times) to achieve stationarity, fitting ARMA(p,q), then reversing the transformation for forecasts—making it practical for real-world time series with trends.