Common Diagnostic Issues and Solutions
You'll learn: How to identify and fix three critical problems that can undermine your time series models: heteroscedasticity, leftover seasonality, and structural breaks in residuals.
Why Residual Diagnostics Matter
After fitting an ARIMA or other time series model, your residuals should look like pure white noise—random, independent, and well-behaved. When they don't, it signals that your model is missing something important. Let's tackle the three most common culprits.
Heteroscedasticity: Changing Variance
The Problem: Your residuals' variance isn't constant over time—sometimes tight, sometimes spread out. Think of a megaphone pattern.
The Solution: Apply a variance-stabilizing transformation to your original data before modeling. Common fixes include log transformation, square root transformation, or Box-Cox transformation. Then refit your model.
Remaining Seasonality
The Problem: ACF plots of your residuals show periodic spikes at seasonal lags—your model didn't capture all the seasonal patterns.
The Solution: Add or increase the seasonal differencing order, or switch to a seasonal ARIMA model with explicit seasonal terms (P, D, Q). Review your seasonal decomposition to ensure you've identified the correct period.
Structural Breaks
The Problem: Your model works well for part of the series but suddenly fails—residuals show persistent patterns after a specific time point. This happens when the underlying data-generating process fundamentally changed (policy shift, market disruption).
The Solution: Split your data at the break point and model each segment separately, or use intervention analysis to explicitly model the break with dummy variables. Alternatively, consider models that allow time-varying parameters.
Key Takeaway: Residual diagnostics aren't just box-checking—they tell you exactly what your model is missing. Heteroscedasticity needs transformation, remaining patterns need better model specification, and structural breaks need segmentation or intervention modeling.