Seasonal ARIMA (SARIMA)
What you'll learn: How to extend ARIMA models to capture repeating seasonal patterns in time series data using additional seasonal components.
The Core Idea
While regular ARIMA models handle trends and short-term dependencies, many real-world time series exhibit seasonal patterns that repeat at fixed intervals—like monthly sales spikes every December or weekly traffic patterns. SARIMA extends ARIMA by adding a second set of autoregressive (AR) and moving average (MA) terms that specifically model these seasonal cycles.
SARIMA Notation: (p,d,q)(P,D,Q)ₛ
SARIMA has two sets of parameters:
- Non-seasonal part (p,d,q): Same as regular ARIMA—autoregressive order, differencing, and moving average order
- Seasonal part (P,D,Q)ₛ: Seasonal AR order, seasonal differencing, seasonal MA order, with period
s
The seasonal lag s defines the cycle length (e.g., 12 for monthly data with yearly seasonality, 7 for daily data with weekly patterns).
How It Works
Think of SARIMA as two interwoven models working together:
- Seasonal differencing (D): Removes the seasonal trend by subtracting values from
speriods ago - Seasonal AR terms (P): Relate current values to values at seasonal lags (e.g., this January to last January)
- Seasonal MA terms (Q): Model seasonal shock patterns that repeat
For example, SARIMA(1,1,1)(1,1,1)₁₂ on monthly sales data would difference the series once normally, once seasonally (12 months apart), then model both immediate dependencies and year-over-year dependencies.
When to Use SARIMA
If your ACF plot shows significant spikes at seasonal lags (12, 24, 36... for monthly yearly patterns), or if seasonal decomposition reveals strong periodic components, SARIMA is your tool. It captures both the short-term dynamics and the repeating seasonal structure in one unified model.
Key Takeaway: SARIMA extends ARIMA with seasonal (P,D,Q)ₛ parameters to model recurring patterns at fixed intervals, combining trend and seasonal dynamics into a single forecasting framework.