Information Criteria: AIC and BIC
You'll learn how to use AIC and BIC to objectively compare ARIMA models and pick the best one.
Why Model Selection Matters
After estimating several ARIMA models with different (p, d, q) combinations, you need a systematic way to choose the best specification. You can't just pick the most complex model—it might overfit. Information criteria give you a principled, numerical way to balance model fit against complexity.
What Are AIC and BIC?
AIC (Akaike Information Criterion) and BIC (Bayesian Information Criterion) are scores that penalize models for using too many parameters while rewarding good fit to the data.
- Lower is better: The model with the smallest AIC or BIC is preferred
- They penalize complexity: Adding unnecessary parameters increases the score
- BIC penalizes complexity more harshly than AIC, especially with larger datasets
Think of it like shopping: you want the best product (fit) for the lowest price (complexity). AIC and BIC calculate that "value score" automatically.
How to Use Them
- Estimate multiple ARIMA models (e.g., ARIMA(1,1,1), ARIMA(2,1,0), ARIMA(0,1,2))
- Calculate AIC and BIC for each model
- Compare the values—the model with the lowest AIC or BIC wins
- If AIC and BIC disagree, BIC typically favors the simpler model
Most time series software reports both criteria automatically after fitting. You simply compare the numbers across candidate models.
AIC vs BIC: Which to Use?
- AIC tends to select slightly more complex models
- BIC leans toward simpler, more parsimonious models
- In practice, try both and see if they agree; disagreement often signals you're near the optimal complexity
Key Takeaway: AIC and BIC provide objective, numerical criteria to select the ARIMA model that best balances fit and simplicity—choose the model with the lowest score.