Model Selection and Diagnostics
You'll learn how to choose the best time series model and verify it's working correctly using diagnostic tools.
Why Model Selection Matters
You've learned about MA, AR, ARMA, ARIMA, SARIMA, and exponential smoothing—but which one should you use for your data? Model selection is about systematically choosing the model that best captures your time series patterns without overfitting.
Think of it like buying shoes: you need the right fit (not too tight, not too loose) and you check by walking around to make sure they work properly.
The Diagnostic Toolkit
ACF and PACF Plots
These plots (which you learned earlier) help identify model orders. Sharp cutoffs in PACF suggest AR terms; sharp cutoffs in ACF suggest MA terms. They're your first visual guide to model structure.
Information Criteria: AIC and BIC
AIC (Akaike Information Criterion) and BIC (Bayesian Information Criterion) balance model fit against complexity. Lower values are better. They penalize adding unnecessary parameters, preventing overfitting.
- AIC is less strict about complexity
- BIC penalizes extra parameters more heavily
- Compare multiple candidate models and pick the one with minimum AIC/BIC
Residual Analysis
After fitting a model, examine the residuals (actual - predicted values). Good residuals should look like white noise:
- No autocorrelation (ACF plot should show no significant spikes)
- Normally distributed (histogram or Q-Q plot)
- Constant variance over time (no patterns in residuals vs time plot)
If residuals show patterns, your model missed something important in the data.
The Workflow
- Use ACF/PACF to guess initial model orders
- Fit several candidate models
- Compare AIC/BIC scores
- Check residuals of the best model
- If residuals aren't white noise, refine and repeat
Key Takeaway: Model selection combines visual diagnostics (ACF/PACF), quantitative metrics (AIC/BIC), and residual analysis to find a model that captures patterns without overfitting—residuals should resemble white noise.