Course contentsShow
← Data Science
Lesson 798 of 2,14520. Time Series ModelsFree lesson

SARIMA Model Selection

Choosing seasonal and non-seasonal orders using information criteria, ACF/PACF, and grid search approaches.

SARIMA Model Selection

What you'll learn: You'll learn how to systematically choose the best seasonal and non-seasonal parameters for your SARIMA model using multiple complementary approaches.

The Core Challenge

Selecting a SARIMA model means picking seven orders: non-seasonal (p, d, q) and seasonal (P, D, Q, s). Unlike simpler ARIMA, you now have both regular and seasonal patterns to capture. Choosing poorly means underfitting (missing patterns) or overfitting (modeling noise). The key is combining statistical evidence from multiple sources to make an informed choice.

Three Complementary Approaches

1. Information Criteria (AIC/BIC) Fit multiple candidate models and compare their AIC or BIC values. Lower values indicate better trade-offs between fit quality and complexity. BIC penalizes complexity more heavily, helping you avoid overfitting.

2. ACF/PACF Analysis After seasonal differencing, examine both plots:

  • ACF suggests the seasonal MA order (Q) by seasonal lag spikes
  • PACF suggests the seasonal AR order (P) by seasonal lag cutoffs
  • Non-seasonal patterns at early lags guide (p, q) selection

3. Grid Search Test a range of parameter combinations systematically—for example, p ∈ {0,1,2}, P ∈ {0,1,2}—and select the model with the lowest AIC/BIC. This ensures you don't miss better configurations your initial ACF/PACF analysis might have overlooked.

The Workflow

Start with seasonal differencing to remove seasonality. Check ACF/PACF for initial guesses. Then run a grid search over plausible ranges, comparing AIC/BIC. Finally, validate your top candidate with residual diagnostics to confirm it captures all patterns without overfitting.

Key Takeaway: SARIMA model selection combines ACF/PACF intuition with systematic grid search and information criteria comparison to find the optimal balance between capturing seasonal patterns and avoiding overfitting.